Nobel Prize in Physics goes to Francis Halzen
The page announces the 2026 Nobel Prize in Physics, noting that the award announcements will occur from 5 to 12 October and can be viewed live on NobelPrize.org. It includes two images: one depicting Francis Halzen and another prompting viewers to watch the live announcements.
Beam: Reflection's 501B open-weight model
Beam is Reflection’s first open‑weight model, a 501 B‑parameter Mixture‑of‑Experts system with 23 B active parameters, optimized for coding, reasoning, and agentic tasks.
- RL environment pipeline: ~1 M synthetic and sourced environments across software engineering, terminal use, competitive coding, STEM, web search, and tool use; iterative filtering for difficulty and quality; RL testing to refine tasks, preventing capability plateaus.
- RL infrastructure: Asynchronous platform sustaining ~110 K concurrent rollouts, flexible inference‑to‑training GPU ratios (3.9:1–5.4:1), median 12 s model‑update propagation, 71 inference incidents handled with <0.02 % GPU‑minute loss, up to 170 K sandboxes, >1 B sandbox creations, 99.99 % trainer batch utilization, per‑token observability for reward integrity.
- Data curation: 23.8 T high‑quality tokens from web, code, and STEM sources; custom quality classifiers and tiered weighting; ~95 % of raw tokens removed, yet 1.8 T high‑quality tokens retained (including 87 % of curated web‑code tokens); language‑specific code filters, PDF OCR pipeline, controlled repetition with fuzzy deduplication.
- Pretraining stack: Completed in <4 weeks on 6,144 NVIDIA GPUs using an in‑house topology‑aware Kubernetes scheduler, node‑lifecycle monitoring, and silent‑data‑corruption detection; nine semi‑automatic rewinds; 92.3 % goodput.
Results show frontier‑level inference efficiency, stable training despite weight staleness, Pareto‑optimal loss on decontaminated validation sets, near‑uniform MoE expert utilization, bounded residual‑stream RMS, and predictable alignment‑RL reward improvements.
Comments show a mixed reaction to Reflection’s Beam model. While the open‑weight release is praised as a noteworthy milestone that could increase competition with Chinese labs, many express skepticism over the lack of publicly available weights, limited technical details, and marketing‑focused claims. Repeated concerns focus on the model’s practical viability, hardware requirements, benchmark transparency, and safety alignment. Comparisons to other large open models highlight perceived underperformance, and there is broader unease about the organization’s sustainability, funding, and long‑term commitment. Overall sentiment balances cautious optimism with critical demand for clearer evidence and accessibility.
Accountability mechanisms can be joyful (2024)
The author describes a music‑based accountability system used with a partner: each person sends a song to the other after completing a daily task (concussion exercises or a workout). This approach provides a low‑stakes, positive reinforcement—failure merely prevents sharing music, while success delivers a pleasant, personal reward for both participants. The method contrasts with traditional “accountability bets,” where unmet tasks incur a concrete penalty (e.g., monetary loss or forfeiting a habit). While bets can be effective for large, high‑stakes projects, they can induce stress, shift intrinsic motivation to extrinsic fear, and be uncomfortable to negotiate. The music system instead “waters the carrot,” leveraging relational joy and emotional support, especially suited for recurring small tasks, though it may be insufficient for more demanding goals.
The comments express enthusiasm for using personalized, game‑like interfaces to boost motivation and celebrate small achievements, highlighting how custom visual trackers can make goal‑setting more enjoyable. They also recognize that such external cues—jingles, memes, or playful competition—can be effective shortcuts for habit formation. At the same time, there is a broader reflection that reliance on these incentives points to a societal challenge in fostering intrinsic motivation for healthy choices, suggesting a need to address underlying behavioral patterns.
Find the flattest route between any two points in SF
The “Flatten SF” tool computes routes in the browser across more than 160 000 San Francisco street segments using USGS 1‑m lidar elevation data combined with the Overture/OpenStreetMap street graph. A slider lets users explore the full trade‑off between distance and climbing: it enumerates every Pareto‑optimal route from the shortest path to the flattest walkable route, assuming a cost of 200 ft of walking for each foot of ascent; routes exceeding this cost are excluded. Moving the slider right never reduces distance or adds elevation gain. Climbing is measured as cumulative ascent, not net elevation change. Stairways are included for pedestrian routes but omitted for bicycle routes. Light lines display alternative routes along the path. Place search (intersections, venues, addresses) operates offline using data embedded in the page. Source code, data, and complete analysis are provided.
Comments express overall appreciation for a Bay Area bike‑routing tool that incorporates detailed elevation data, noting its usefulness for finding flatter paths and its appealing visual interface. Users frequently request additional features such as explicit distance‑and‑elevation metrics, options for steepest routes, drive modes, and clearer street‑name toggles. Several criticisms point to inaccurate routing in specific locations, occasional discontinuities, and limited elevation resolution compared with other services. Comparisons to existing map platforms highlight perceived shortcomings in hill handling, while suggestions include integrating alternative engines and improving data attribution.
Dust: Pretraining Transformers Without Backpropagation
Dust introduces a zeroth‑order optimizer for pretraining transformers that perturbs activations (“node perturbation”) independently at each token, treating each token as a member of a virtual population evaluated in a single forward pass. The reward‑weighted perturbations yield an estimated output error whose outer product with the layer input provides a weight gradient, mirroring backpropagation’s update but without differentiability requirements. Compared to weight‑space evolution strategies such as EGGROLL, Dust is 10³–10⁴ times more compute‑efficient from 1 M tokens onward, because activation‑space perturbations avoid materializing full weight copies. Empirical studies on GPT‑style models (8 layers, width 512) pretrained on FineWeb show that Dust matches backpropagation at modest token budgets and surpasses it at larger budgets (10–20 M tokens) when population size is increased, with gradient cosine similarity to backprop improving as population grows. Larger models (e.g., 243 M parameters) demonstrate greater population efficiency than much smaller counterparts, contradicting conventional expectations about over‑parameterization. Dust thus provides a scalable, backprop‑free alternative for transformer pretraining, though further work is needed for compute‑parity with standard training.
Comments show a prevailing view that gradient‑based methods remain superior for typical neural‑network training because objectives are smooth and gradients efficiently guide optimization, while derivative‑free approaches are seen as costly, slower, and offering limited advantage except when gradients are unavailable, such as with simulators. Skepticism surrounds claimed benefits of the new zeroth‑order technique, with calls for concrete measurements of runtime, energy, memory, and quality at comparable loss. Interest exists in hybrid or specialized optimizers and in exploring evolutionary ideas, but overall consensus holds that backpropagation remains the most practical standard.
Friendship ended with Deno, now Node is my best friend
The author migrated a SvelteKit client project and a static‑site generator from Deno to Node v26.10.0, noting that modern Node now supports most ECMAScript features without require() and offers improved APIs (e.g., node:fs replaces Deno’s file‑system API, and Deno.serve is swapped for Hono’s node adapter built on node:http). A simple import change from @std/path to node:path enabled a 15 % build‑time improvement.
Package management is handled with pnpm, aliased to npm/npx, and configured via pnpm‑workspace.yaml (minimumReleaseAge 1440 minutes, trustPolicy no‑downgrade) to limit automatic updates and mitigate supply‑chain risks. The author disables post‑install scripts and adjusts trust policies for self‑hosted Forgejo packages.
Node’s handling of TypeScript remains restrictive: it rejects *.ts files inside node_modules, requiring pre‑compilation (e.g., with tsd). The post‑migration setup works despite these constraints.
The piece also outlines Deno’s perceived decline—issues with ZSH integration, JSR rate limiting, and concurrency bugs—leading to its removal (brew uninstall deno). Overall, the transition demonstrates that contemporary Node can replace Deno with minimal code changes and performance gains.
The discussion highlights strong appreciation for Deno’s integrated test runner, linter, type‑checking and security model, with many valuing its minimal configuration and consistent TypeScript version. Concerns are voiced about recent layoffs, unclear roadmap, and ecosystem fragmentation that may limit adoption. Node is regarded as mature and reliable, though occasionally criticized for slower performance and heavier setup. Bun generates interest for speed and modern features but faces criticism over compatibility bugs. Overall sentiment is mixed: enthusiasm for Deno’s design is tempered by worries about its long‑term viability, while Node remains the safe default and Bun is seen as an emerging alternative.
Opus 5.5 agents discover two room-temperature magnetic semiconductor candidates
The article discusses Luttinger‑compensated (LC) antiferromagnetic semiconductors, which combine zero net magnetization with spin‑sorted electronic states—features desirable for spintronic memory. Two candidate materials are presented:
-
YBaMnFeO₅ – a newly designed oxide predicted by density‑functional theory (PBE+U and HSE06) to be a semiconductor with a 2.35 eV band gap, spin‑sorting windows of ~1.0 eV (holes) and ~1.4 eV (electrons), and magnetic order up to ~420 K (≈490 K after calibration). Its LC behavior requires a perfect Mn/Fe checkerboard; simulations show this ordering breaks down near 950 K, making synthesis of the ordered phase challenging.
-
KV[Cr(CN)₆] – a Prussian‑blue‑type compound synthesized in 1999. Calculations predict a 2.1 eV band gap, spin‑sorted windows of 2.6 eV (holes) and 1.6 eV (electrons), and magnetic ordering up to 376 K. The structure intrinsically locks Cr and V on distinct cyanide sites, preserving LC order. Experimental samples contain pore water, giving a small residual moment; HSE06 suggests spin sorting survives, while PBE+U predicts a reduced hole window.
Both materials exhibit zero net spin, sizable band gaps, and spin windows far exceeding thermal energy (≈26 meV), indicating potential for room‑temperature spintronic devices if the LC order can be realized experimentally.
The comments show widespread skepticism toward the headline claim of “room‑temperature magnetic semiconductors,” emphasizing that the reported results are merely computational predictions that require experimental verification and that similar materials have been known previously. Many criticize the sensational wording and marketing tactics of the AI‑driven research group, while acknowledging that using large‑language‑model agents to automate DFT screening can accelerate hypothesis generation. Opinions also note the need for clearer explanation of magnetic ordering, caution about over‑reliance on AI without human expertise, and modest optimism that transparent code and data sharing could make the approach useful once laboratory synthesis validates the candidates.
Web Search API
The Web Search API is now in beta, enabling AI agents and applications to perform live Internet searches and base responses on current data rather than model training limits. At launch, three providers are supported—Ceramic.ai, Exa, and Linkup—each offering zero‑data‑retention for Cloudflare‑routed requests and adhering to Cloudflare’s verified bot‑crawling standards. The API operates through AI Gateway, so search calls appear in gateway logs and are billed to AI Gateway credits at each provider’s listed API price with no additional markup; users may also supply their own provider API keys. Calls can be made via REST (e.g., curl with JSON payload specifying query, provider, limit, and gateway ID) or through Cloudflare Workers using the AI binding. Documentation for initial setup is available under “How to use Web Search API.”
Comments emphasize strong concern over search‑API restrictions that prohibit storing or resyndicating results, viewing this as a major limitation for agent systems. Pricing is repeatedly described as high and uneven across providers, prompting interest in cheaper or free alternatives such as self‑hosted solutions (SearXNG, Firecrawl) and smaller APIs (Kagi, Serper.dev). Opinions on Cloudflare’s Web Search proxy are split: some appreciate unified access and potential latency gains, while many criticize unclear data‑retention guarantees, perceived unnecessary indirection, and possible antitrust issues. Overall, users seek transparent terms, consistent retention policies, and cost‑effective, high‑quality search options.
Resurrecting iChat Audio and Video Conferencing
Add a DNS entry for configuration.apple.com pointing to a reachable SNATMAP server (e.g., 157.230.2.213) on Macs running OS X Leopard or Snow Leopard. iChat AV originally obtains its public IP via an HTTP request to http://configuration.apple.com/configurations/macosx/ichat/1/snatmap.txt, then sends a UDP request to the returned SNATMAP server (port 5678) which replies with the client’s external IP and port. Modern Apple servers no longer respond, causing iChat to exchange private LAN addresses and fail NAT traversal. The author reverse‑engineered the SNATMAP protocol (UDP, 16‑byte request: type 1, nonce, unused local IP/port; 16‑byte response: type 2, same nonce, external IP, external port) and implemented a minimal Python server. By hosting the SNATMAP file on a custom web server and ensuring source‑port‑preserving NAT on routers, iChat can re‑establish peer‑to‑peer RTP audio/video calls, even with non‑Apple webcams. The solution works with typical NAT routers and BSD‑based routers after enabling source‑port preservation.
Comments express nostalgia for iChat AV’s early video‑conferencing capabilities, recalling its development, hardware integration and the appeal of the iSight camera as both a functional device and a collectible item. Users contrast this with frustration toward recent FaceTime updates, describing the interface as increasingly confusing and prone to accidental calls. Technical observations note the importance of NAT handling and the lack of TURN support in iChat, while overall sentiment blends fond appreciation for the legacy technology with criticism of newer implementations.
Example.com just launched the biggest redesign in decades
Example.com is a reserved domain used for documentation examples. Since its 1999 launch, the site has undergone sporadic redesigns and infrastructure changes:
- Early versions (2002) displayed a table‑based list of reserved names on Apache 1.3.22; messaging was added in March 2002 and a RFC 2606 citation in Feb 2003 (Apache 1.3.27 on Red Hat).
- July 2010 added example.edu; July 2013 removed a redirect to IANA, moved hosting to the EdgeCast CDN, and introduced a rounded‑card layout.
- Oct 2019 updated font stacks and refined copy; Jan 2025 the server header vanished after EdgeCast’s transition to Edgio (later acquired by Akamai).
- Oct 2025 redesign minified HTML, altered link text, and dropped the card; Dec 2025 migrated to Cloudflare (CF‑RAY header); Jun 2026 added a data‑URI favicon to eliminate extra requests.
- Sep 28 2026 overhaul adds JavaScript‑driven multi‑language rotation (English, Arabic, Chinese, French, Russian, Spanish) with a 0.4 s opacity transition; each character is wrapped in a span with incremental transition‑delay, and an SVG book icon is inserted.
IANA notes the changes aim to reduce bandwidth and discourage reliance on example.com for testing or monitoring.
Comments converge on criticism of the recent redesign, noting that heavy compression, minified HTML and added JavaScript increase page size, break fragile automated tests, and waste bandwidth for a domain meant to be a minimal, static reference. Users question the reliance on Cloudflare, the lack of a simple fallback, and the decision‑making process at IANA, while some acknowledge the move toward multilingual support as understandable. Overall sentiment is that the changes add unnecessary complexity to a service intended for lightweight, reliable testing and documentation.
The response conveys strong enthusiasm for the IceCube neutrino detector and the recent Nobel Prize, highlighting the project's bold, sci‑fi‑like ambition of installing sensors in Antarctic ice to capture elusive particles. It emphasizes the rarity and prestige of Nobel recognition, noting the large number of living laureates and the significance of a single physicist receiving the award. The tone is celebratory and appreciative, mentioning related discussions and visual materials while expressing relief that the prize honors fundamental physics rather than AI.