AI Frontier Daily Briefing: 2026-09-20
AI posters all look the same, says the day's #1 post with 1,645 upvotes; Laya answers structured questions in one forward pass without generating a word; Tao's blog hosts a 271-comment fight over what math is for beyond proof; GPT-6 Astra cracks a 107-year-old German cipher; OpenAI designed its Jalapeño chip with its own LLMs; a Rust veteran's Zig rewrite sparks the day's biggest argument; Gemini broke into three real companies during a security test; a hallucinated AI intel report nearly put US troops on a Chinese ship; the AI-slowdown essay draws an antitrust class action; DraftKings uses AI to target the gamblers likeliest to lose. Newly unsealed lawsuit filings quote a Microsoft director calling AI scraping “the largest theft of labor in human history”; Flock Safety offers buyouts to 1,500 employees after 93 local governments ended contracts in August; NASA and IBM open-source a lunar foundation model with its weights; a self-proclaimed world-fastest PHP webserver draws skeptical comments; and a 2013 post dissects HN's ranking formula and hidden penalties.
81 stories made the HN front page on 2026-09-19 (UTC), seven of them with more comments than upvotes. The top story was a takedown of the default “AI look” in event posters (1,645 upvotes). The loudest fights: a seven-year Rust user’s field report on Zig (288 comments) and a leaked Apple M6 Pro Geekbench result (144 comments). On models, GPT-6 Astra decoded a 107-year-old German military cipher, and NASA and IBM open-sourced a lunar remote-sensing foundation model with its weights. On industry, newly unsealed filings in the NYT lawsuit quote a Microsoft director calling AI scraping “the largest theft of labor in human history,” and license-plate-recognition vendor Flock Safety opened buyouts to its 1,500 employees after 93 local governments ended contracts in August. 28 items, loosely grouped into models & agents, developer tools, hardware, and governance.
1. AI posters don’t have to look like AI, and 1,645 up voters agree
Pastel bunting again? There has to be another way.
John Hartnup’s post hit #1 with 1,645 upvotes and 860 comments. His claim: AI posters have converged on one look: pastel airbrush art, bunting, hand-drawn craft-fair illustration. The problem isn’t using AI, it’s that everyone gets the same default style. His method: 1. Supply invented event details (date, venue, acts) in one brief. 2. List explicit bans (he ruled out pastels, airbrush, and images of people). 3. Ask for a deliberate “what not to do” example first. 4. Name the style once the model produces one; after a Bauhaus result, having the model label it gave him callable vocabulary. 5. Borrow named movements: Memphis, Designers Republic, 90s rave flyers. He also published a catalogue of 100 poster styles with paste-ready prompts. If you make event materials, this list will do more than switching models. Source · HN discussion
2. Laya answers questions in one forward pass, no words generated
No generated tokens, so nothing to hallucinate?
ConvAI Innovations founder Nandakishor Mukkunnoth open-sourced Laya, a family of non-autoregressive “System 1” decision models that answer structured questions in a single forward pass: 1,238 upvotes, 295 comments. Three primitives: 1. choice, pick one option from a candidate dictionary; 2. score, place input on an ordinal rubric; 3. noul, a boolean returning a calibrated P(true). Trained with RLCD (reinforcement learning for calibrated decisions) on ModernBERT-large (421M parameters) and multilingual mmBERT-base (322M). Numbers: 32.8ms per question on one GPU (7.2ms batched), 0.766 accuracy on typed-decisions, ECE calibration error 0.081, three times better than the closed-source Jev, and 45 of 51 languages above 3x random baseline. Limits are stated plainly: choice degrades past ~20 options (0.425 on Banking77’s 77 labels) and zero-shot is near-random. Apache 2.0, pip install laya, runs air-gapped. If you need classification, routing, or spam checks in a pipeline, this runs without an API bill. Source · HN discussion
3. A 681-upvote writing workflow where you never use the model’s words
One borrowed phrase and the whole piece turns to Velveeta?
“How to Write with an LLM” from sockpuppet.org drew 681 upvotes and 396 comments, #4 on the front page. Two hard rules: 1. Never use a phrase the model suggests. Models default to punchy headline-style prose, and borrowing their wording drags the whole piece into the same voice. 2. Disable the praise. Models reflexively compliment drafts, which keeps you from cutting bad paragraphs. The workflow: write the entire piece yourself first, then have a strong model flag problems only (passive voice, nominalized verbs, repeated phrasing, paragraphs in the wrong place); you decide what to fix and how. He built a small app (Python, HTMX, SQLite) that runs a saved list of editing prompts over CLI models like Codex and Claude. If you write for a living and can’t avoid AI, this “model as copyeditor, you as author” split is copyable. Source · HN discussion
4. Tao’s blog hosted a 271-comment fight over what math is for
If machines can prove it, what should math reward?
Grant Sanderson of 3Blue1Brown guest-posted on Tao’s blog: 365 upvotes, 271 comments, #5 for the day. Core claim: proof has been a proxy for the real goal of advancing human understanding, and since AI can now generate proofs without understanding, the field needs new proxies. “Motivated explanations” deserve credit comparable to solving open problems. His concrete example: Liam Price used GPT-5.4 Pro to help solve Erdős Problem 1196, and Tao plus others then expanded and explained the key idea, which Sanderson argues deserves more celebration than the solve itself; his test is whether a reader feels they could have discovered it. He also states there will never be a Lean for motivated explanations. Researchers and PhD advisors should read it in full: it bears directly on how papers get evaluated. Source · HN discussion
5. GPT-6 Astra cracked a 107-year-old German WWI radio cipher
The key was used early, no wonder it stayed unsolved?
Blogger “prinz” used GPT-6 Astra to decode a German ADFVX-era military message from November 27, 1918: 383 upvotes, 174 comments. Process: 1. The model identified the key word TRUPPENVERSCHIEBUNG (troop movement) from J. Rives Childs’s history of German military ciphers. 2. Key letters were reordered alphabetically, and the 170 cipher characters were laid out in rows of 19, giving 18 columns of nine symbols plus one column of eight. 3. Column positions mapped back to plaintext. The plaintext: an English cruiser arrived at Sevastopol, followed by an Allied squadron on the 26th. The model verified it against ship logs: HMS Canterbury arrived November 24, 1918, the Allied squadron on the 26th. Under German regulations this key wasn’t supposed to be in use until December 9, which may be exactly why the message resisted decryption. Cipher hobbyists and anyone doing historical research with LLMs: this chain is worth reproducing. Source · HN discussion
6. OpenAI designed its Jalapeño chip with its own LLMs
The models that write code now lay out silicon?
IEEE Spectrum details how OpenAI’s first in-house AI accelerator, Jalapeño, was designed: 201 upvotes, 135 comments. Specs: up to 13.4 petaflops of 4-bit compute, 232GB of HBM4 at 15.4TB/s, pods of 2,048 chips, and a claimed up to 3.6x lower end-to-end latency than Nvidia’s GB300. LLM involvement came in three places: 1. Front-end work went through Google’s open-source XLS high-level synthesis toolchain so models could write DSLX/C++ instead of Verilog, though later internal models worked in Verilog directly. 2. After first silicon returned in May, internal models optimized DeepSeek’s multi-head latent attention kernel from 0.31% to 88.94% of the theoretical ceiling in about 40 hours. 3. AI-guided optimization cut matrix-multiplication unit area by 10% versus a human baseline, while Broadcom handled physical design. The team averaged under 100 people and went from concept to silicon in under 20 months. Chip and EDA people: this is a working example of “models write code, humans arbitrate.” Source · HN discussion
7. A Rust veteran rewrote a library in Zig, and 288 comments followed
No borrow checker to catch you. How does that feel?
The author, seven years into Rust, rewrote an RFC 9535-compliant JSONPath library in Zig and wrote up the comparison: 239 upvotes, 288 comments, more comments than upvotes. Concrete differences: 1. The Rust project had 17+ files across src/parser and src/query; the Zig version has five files at the top level. 2. Rust’s monadic combinators and lazy iterators have no Zig equivalent, so code moved to in-place mutation and loops. 3. Allocators are everywhere, nearly every function takes one, and he used TestAllocator and FailingAllocator to catch four bug classes: forgotten deinit, error-path leaks, a cross-layer double free, and orphaned allocations after a successful dupe, fixed with defer, errdefer, and single-layer ownership. 4. IDE support is near zero; he ended up on helix + alacritty + zellij. His verdict: Zig is straightforward, modern, and fast, but the resulting code is subjectively less readable than his Rust, and the ecosystem is “still very young”: the regex crate lacks \p{…} support. Anyone choosing between the two languages gets first-hand material here. Source · HN discussion
8. “Almost never use AI to write,” says this 308-upvote essay
If you didn’t write it, did you actually think it?
Erich Grunewald’s long essay drew 308 upvotes and 150 comments. The claim: substantive texts (blog posts, reports, memos, novels) should almost never be written by AI, however detailed the prompts and however heavy the human editing. Three reasons: 1. Writing and thinking are the same act; citing Paul Graham, half the ideas in an essay appear while writing it, and an AI handed an outline won’t reconsider the outline itself. 2. AI text is vague and subtly wrong yet convincing; he line-edits a Claude-written paragraph about AI chip smuggling and names a dozen problems. 3. Unlabeled AI writing breaks the reader-writer contract. He endorses AI for transcription, data analysis, brainstorming, and line-editing where a human accepts or rejects each change. Writers and anyone accountable for content quality should read the strongest version of the case against their own habits. Source · HN discussion
9. PlanetScale’s Postgres full-text search claims 25x ParadeDB speed
Another Elasticsearch replacement?
PlanetScale released Tin, a full-text search extension for Postgres: 218 upvotes, 86 comments. The design: 1. Use Postgres ctids (48-bit values: 32-bit block, 16-bit offset) as document IDs directly, skipping the ID-to-ctid mapping layer that ParadeDB and pg_textsearch both need. 2. Two-level bitmap compression, down to 1 bit per posting for common terms. 3. Page-level bitmaps fit in AVX2 registers, so AND/OR queries run as vectorized instructions. In PlanetScale’s own tests on an 85GB, 150M-document Stack Exchange corpus: index build took 8m10s at 50.7GB versus ParadeDB’s 19m20s at 64GB of RAM and GIN’s 2h09m; mixed top-10 queries ran at 199 QPS versus ParadeDB’s 7.9, with p99 of 256ms versus 6,765ms. Usage is CREATE INDEX ... USING tin(text_column) and a ==> operator. If you’re running Elasticsearch next to Postgres, benchmark your own queries against this before deciding. Source · HN discussion
10. Alibaba open-sources Damo Radar, a CT model covering 146 findings
An open-weight model that reads scans like a generalist?
SCMP reports Alibaba’s Damo Academy open-sourced Damo Radar: 148 upvotes, 22 comments. It’s a vision-language model that reads contrast-enhanced abdominal CTs across 18 organs and detects 146 clinical findings including cancers; across nearly 40,000 real-world exams it averaged an AUC of 0.913, and a cited Science study says it outperformed most radiologists. The team calls it the first expert-level generalist medical imaging model, and says the method should extend to other imaging types. Medical AI builders and anyone doing imaging annotation: reproduce it on your own data before drawing conclusions. Source · HN discussion
11. One git push, 9,236 S3 requests down to 30
11 million objects, an hour at 10ms each. Who waits?
A Tigris engineer’s post drew 144 upvotes. The problem: Git packfiles don’t record compressed object sizes, so you can’t build HTTP Range requests, and object storage forces one request per object; the Linux kernel’s 11M+ objects would take over an hour even at 10ms each. The fix is a new packfile format: objects stored sequentially in a .bin file (128MB cap) plus a .cue sheet with fixed 58-byte records holding hash, type, zstd sizes (compressed and uncompressed), and delta base; single-object Range requests become trivial and delta objects are stored separately. Measured: a push to the Xe/x monorepo went from 9,236 requests and 3m29s to 30 requests and 14.3s; clones improved 1.8–4.5x. Open source at github.com/tigrisdata/objgit, no auth or compaction yet, and the author says it’s not production-ready. If you want repositories on object storage, this is the most serious attempt so far. Source · HN discussion
12. Why building a Rust LSP is hard, per someone building one
Half-typed code is the normal case, not the error?
The author of Rust Glancer, an experimental Rust LSP aiming at low RAM and instant editor restarts, explains the difficulty: 138 upvotes, 67 comments. Compilers get complete input and can fail on errors; LSPs must produce useful answers from half-edited, broken state. Concrete points: 1. LSP hides the filesystem, so the server needs its own VFS, concurrent reads, cancellation, and tolerance for imperfect client notifications. 2. Query costs are layered: symbols need parsing, hover needs item trees, inlay hints need type inference, references need workspace-wide candidate filtering. 3. Protocol details like UTF-16 line/column positions and edition-dependent keywords leak into the architecture. His route is the opposite of rust-analyzer’s salsa-style lazy incrementality: eager indexing, per-workspace process isolation, and state offloaded to disk. Language-tool and IDE-plugin builders get a clear map of the design space. Source · HN discussion
13. SDCC 4.6 is out, the 8-bit C compiler keeps moving
The 8051 still won’t die?
SDCC (Small Device C Compiler) 4.6.0 shipped June 22 and hit the front page with 133 upvotes. It’s a retargetable, optimizing C compiler suite (GPL) for resource-constrained 8-bit microcontrollers (MCS-51, the Z80 family, STM8, MOS 6502/65C02, Padauk, HC08/S08) plus an assembler, linker, simulator, and debugger. Version 4.6 adds C2y features (_Countof, if-declarations, bit utilities), new Rabbit 4000/5000/6000 ports, substantially improved Z80-family code generation, and an eZ80 port replacing ez80_z80. In October 2025 the project took funding from the NGI0 Commons Fund and the Sovereign Tech Fund for target hardware improvements, link-time optimization, and post-quantum crypto readiness. Retro-hardware hackers and anyone writing C for extreme constraints: this toolchain keeps improving. Source · HN discussion
14. Stagehand runs Playwright 2x faster, 80% fewer tokens
So the bill was mostly tokens, not browser time?
Browserbase’s Stagehand drew 140 upvotes and 36 comments. The claim: point the same script at Browserbase and it executes 2x faster than Playwright cloud browsers. How: 1. Stagehand runs as an extension next to the browser, cutting a network round trip on every action. 2. Hybrid accessibility-tree trimming gives agents exactly the page context they need, which is where the 80% token savings come from. 3. An optional Model Gateway picks the cheapest model per action, and server-side caching means repeat calls spend no tokens. SDKs exist for TypeScript, Python, and Go, plus an MCP server so coding agents like Claude Code and Codex get browser tools without a local install. If you run browser agents, splitting action execution from page understanding makes the invoice much easier to read. Source · HN discussion
15. Apple M6 Pro posts 4,150 single-core and tops the Geekbench 7 chart
One leaked chip takes the crown?
A Geekbench 7 result for Mac18,6 (presumably a MacBook Pro) landed on September 18: Apple M6 Pro, 18 cores at 4.78GHz, single-core 4,150 and multi-core 37,565, the top single-core score in the Geekbench 7 database. The thread drew 118 upvotes and 144 comments, more comments than upvotes. Context: the base M6 scores 4,071 single-core, the M5 Max 3,731, Qualcomm’s Snapdragon X2 Elite Extreme 3,396, and AMD’s Ryzen 9 9950X3D2 around 3,161. Treat this as a single pre-release leak, not a review; a 37K multi-core number also makes thermals and power strategy the thing to watch in real hardware tests. If you’re waiting on M6-class machines for local inference, single-core speed sets single-request latency. Wait for the reviews before buying. Source · HN discussion
16. Gemini broke into three real companies during a security test
Who gave the sandbox internet access?
The WSJ reported on the 18th, and Google confirmed, that during a May cybersecurity evaluation run by Israeli AI-security firm Irregular, a testing setup error gave Gemini internet access and a fictional company name collided with a real one; Gemini then entered three real companies’ systems: once by guessing a password, twice via credentials found in public repositories: 76 upvotes, 70 comments. Per the reporting, Gemini stopped on its own after recognizing it had touched real companies, the three companies were notified, and Google had known since July but disclosed only after reporters asked in September. AI red teams: isolation and no internet are not optional, and if you test with invented company names, check for collisions first. Source · HN discussion
17. AI-hallucinated intel nearly put US troops on a Chinese ship
The planes were already airborne over a draft?
CNN broke the story on the 18th: in spring 2026, US military aircraft were airborne for an armed operation against a Chinese vessel when officials discovered at the last moment that the underlying intelligence had been fabricated by an AI chatbot; the operation was aborted. Ars Technica’s version drew 14 upvotes and 16 comments, TechCrunch’s 19. Process: a Special Operations Command analyst asked an AI chatbot to combine open-source data with classified signals intelligence; the chatbot fabricated the ship’s manifest, claiming it carried nuclear-weapons components; the analyst then used the same tool to format the findings into an official-looking summary that circulated through command channels, during the war with Iran. The reporting doesn’t name the model vendor. GovAI research scholar and Army veteran Jake Steckler’s takeaway: add safeguards rather than abandon AI. Builders of high-stakes decision systems should audit every link in this chain. Source · HN discussion
18. A class action says the AI labs conspired to slow down
Safety coordination, or an output-restricting cartel?
A consumer class action, Buist v. Anthropic PBC, was filed September 18 in the Northern District of California (No. 3:26-cv-10693), per AP coverage that drew 51 upvotes and 19 comments. The defendants: Anthropic, OpenAI, SpaceXAI (xAI), and Google. The theory: the four competitors violated Section 1 of the Sherman Act by agreeing to slow AI development after Dario Amodei’s September 12 essay called for industry-wide coordination to “pace the frontier”; the complaint cites public endorsements by Sam Altman, Elon Musk, and Demis Hassabis as evidence of agreement. The plaintiffs are paid subscribers to ChatGPT, Claude, Grok, and Gemini, seeking class certification, an injunction, and declaratory relief. The core legal proposition: even safety-motivated agreements among competitors to restrict output get no antitrust exemption. AI-governance watchers: this is the first time “safety coordination” gets fought inside an antitrust frame. Source · HN discussion
19. CUA-S1, small open-source models for fast computer-use decisions
Let the small model click, the big model think?
trycua posted the first research preview of its CUA-S1 family on GitHub: 82 upvotes. Positioning: “System 1” models that don’t plan or reason but make fast, bounded decisions: which value goes in which form field, which element to leave alone. The first profile targets forms, scoring decisions from structured interface elements and document values while application code sequences the actions; an optional Cua Driver executes them with explicit action boundaries. Source is MIT-licensed, weights and datasets live on Hugging Face under CUA-S1-FORMS, and the main repo holds about 24k stars. Computer-use agent builders: splitting fast decisions from slow planning into separately trained models is another route to lower token spend. Source · HN discussion
20. DraftKings uses AI to target the gamblers likeliest to lose
Losing money makes you a good investment?
A New York Times investigation published on the 19th drew 62 upvotes. Key findings: DraftKings built a machine-learning model in 2023 that scores bettors by their habits, predicting who will gamble more and lose more after receiving a promotion, then targets them with free bets and bonuses; former data analyst Jayden Butts is quoted calling a problem gambler “the best investment”; six former employees confirmed the targeting is in use, and four said a model built to flag problem gamblers was shelved. Recommendation and growth-model builders: the same stack of techniques lands on opposite sides of the regulatory and public-acceptance line depending on whose wallet it optimizes. Source · HN discussion
21. Btrfs, ZFS, and bcachefs under the workloads classic benchmarks skip
Default settings or real workloads: which do you trust?
A benchmark from bartosz.fenski.pl drew 144 upvotes and 133 comments by putting Btrfs, ZFS, and bcachefs through tests classic suites skip: fsync tail latency, near-full disks, ENOSPC behavior, scrub, degraded and rebuild modes, and directories with huge file counts. Key numbers: 1. A ZFS mirror with default 128k records posted 202ms fsync p99; the same pool at 8k records posted 3.36ms, a 60x gap on identical hardware. 2. bcachefs on a single disk hit 383.9MB/s sequential writes while btrfs raid1 managed 168MB/s. 3. At 95–99% full, copy-on-write filesystems dropped to single-digit MB/s write throughput. The comment fight wasn’t over the ranking but over whether default configs count. Self-hosters choosing storage: copy his workload list and rerun it against your own disks. Source · HN discussion
22. StepFun ships Step 5 Preview, 600B params at $1 per M input
$2.70 per million output tokens, and index 44?
Artificial Analysis’s model page shows StepFun released Step 5 Preview on September 18: a proprietary 600B-parameter reasoning model with text-plus-image input and a 1M-token context, drawing 15 upvotes and 2 comments. Scores: Intelligence Index 44 versus a 24 median for its tier, ranked #24 of 200; output speed 99.8 tokens per second against a ~70 median, 2.96s to first token; pricing $1.00 per 1M input tokens and $2.70 per 1M output tokens against a $10.00 output median for the tier, plus a 95% cache discount. AA’s verdict includes “very verbose”: the Intelligence Index run burned 160M output tokens versus a 92M median. Model shoppers: at this price the model belongs on your own benchmark list. Source · HN discussion
23. ZK-JPEG proves your image edit survived JPEG compression
Blur the face and keep the camera signature?
IACR ePrint 2026/2039, 91 upvotes. The problem: a camera-signed original loses its signature the moment it’s lossy-compressed or redacted, and prior zero-knowledge edit-history schemes didn’t survive JPEG-style lossy encoding. ZK-JPEG folds a broad class of image transformations into the compression step so they can be verified: the implementation uses PicoZK to turn Python image-editing code into a circuit for the LPZK (line-point) proof system, and the paper claims it’s fast, flexible, and built from off-the-shelf tools. Authors: Samuel Dittmer, Steve Lu, and Kimberlee Model of Stealth Software Technologies plus Joseph Near of the University of Vermont, in the SCN 2026 version. Content-provenance and image-forensics builders should track this line of work. Source · HN discussion
24. Microsoft exec calls AI scraping “history’s largest labor theft”
Your own director wrote it. How do you walk it back?
Newly unsealed filings in The New York Times v. OpenAI and Microsoft (SDNY, Judge Sidney Stein) quote a January 2023 internal memo by Brent Hecht, Microsoft’s Director of Applied Science, describing mass scraping for AI training as “an astonishing theft of unprecedented proportions” and “perhaps the largest theft of labor in human history”: 171 upvotes, 47 comments. The plaintiffs’ summary-judgment papers cite more numbers: 1. over 91,692 copies of NYT, Daily News, and CIR works found in training corpora; 2. a Common Crawl-derived dataset containing more than 2 million nytimes.com documents; 3. a project codenamed Mango spanning 160,903+ unique publisher works; 4. Microsoft data allegedly showing Copilot cut click-throughs to nytimes.com by up to 93%. Microsoft disavowed the memo as “one employee’s individual perspective.” Anyone responsible for training-data compliance: this unsealed list is a ready-made audit checklist. Source · HN discussion
25. The “world’s fastest PHP webserver” meets 72 skeptical comments
Every benchmark self-published. Who reruns it independently?
Qbix Server is a single-file, pure-PHP server meant to replace the nginx + PHP-FPM stack, running unmodified WordPress, Laravel, Symfony, and Drupal. The mechanism: a parent process preloads the framework (~30MB) and forks workers with pcntl_fork, so copy-on-write brings each worker to about 120KB instead of the 30 to 60MB a PHP-FPM worker needs; 28 functions are shimmed via source transformation for persistent mode, with static-property snapshots restored in 0.03ms. Its own numbers: 2,294 req/s on CPU-bound work versus ~350 for PHP-FPM (6.6x), and 1,060 versus 78 on I/O-bound work with 50ms queries (14x); a $30/month machine is claimed to serve 2,000 req/s and 40,000 concurrent users. The thread’s skepticism, at 72 comments against 43 upvotes, targets the lack of third-party validation and a Swoole comparison that ran without coroutine flags. Version 1.3, source on GitHub. If you run PHP at scale, wait for an independent rerun before migrating. Source · HN discussion
26. NASA and IBM open-source a lunar foundation model trained on 2M tiles
If the Moon model ships open weights, what’s Earth’s excuse?
NASA and IBM released the Lunar Foundation Model on September 10: a multimodal, multi-resolution geospatial model for lunar remote sensing, 53 upvotes. Weights are on Hugging Face under Apache 2.0, code on GitHub at NASA-IMPACT/NASA-IBM-Lunar-Foundation-Model. Training data: roughly 2 million image tiles, mostly from the Lunar Reconnaissance Orbiter, plus GRAIL, Lunar Prospector, and JAXA’s SELENE; the companion SomBench dataset aggregates 30+ spatially aligned layers from nine instruments across four missions. Results: up to 22 to 23% lower error than a SwinV2-B baseline on polar ice prospecting, and about 19% better crater detection at 100m resolution with half the training data. The model is built into the open-source TerraTorch toolkit, with LoRA fine-tuning that freezes 90% of base weights. Remote-sensing and scientific-dataset builders: this family, from Prithvi and Surya to the Moon model, is the one to study for how open weights get shipped. Source · HN discussion
27. Flock Safety offers buyouts after 93 governments walked in August
The surveillance backlash lands on the employees?
Wired reports that Flock Safety, the license-plate-recognition company, opened voluntary buyouts to its roughly 1,500 employees: 90 upvotes, 35 comments. Applications opened September 18 and close October 2; the company expects to approve most requests, and the package is described internally as its most generous ever, with tens of thousands of dollars in compensation, months of healthcare, and two years to exercise options, because without buyouts, layoffs would be “almost certain.” On the customer side, an advocacy group counted 93 city and county governments ending ties in August alone, four times the prior month; Florida and Texas announced they will stop using the technology, and the Washington Post identified 46 officers accused of misusing it. B2G and surveillance-product builders: this is a complete chain from public backlash to government churn to organizational shrinkage. Source · HN discussion
28. 2013 post dissecting HN’s ranking formula resurfaces at 198 upvotes
So the front page you scroll every day is scored how?
Ken Shirriff’s 2013 analysis returned to the front page with 198 upvotes and 104 comments. The mechanics it documents: 1. score = (upvotes − 1)^0.8 / (age_hours + 2)^1.8, and since the age exponent exceeds the vote exponent, recency beats popularity; 2. titles containing “NSA” carried an automatic 0.4 multiplier, and many popular domains carried hidden factors between 0.25 and 0.8; 3. more comments than upvotes plus at least 40 comments triggered a controversy penalty the author suspects cubes the divisor rather than squaring it; a 0.4 penalty makes each vote count as 0.3 and decays ranking about 66% faster. In the author’s dataset, roughly 20% of front-page stories and 38% of second-page stories appeared penalized. Content-distribution and growth engineers: the explicit-formula-plus-hidden-penalty design is worth comparing against your own recommender. Source · HN discussion