The hardest working font in Manhattan (2025)
Gorton is a monoline, routing‑type sans‑serif font first supplied with George Gorton Machine Co. pantograph engravers in the early 1900s. Its square‑proportioned, mechanically uniform glyphs—flat R, wavy Q, symmetrical 6/9, and a 3 resembling Cyrillic—were designed for easy tracing and carving into metal, plastic or wood, making the lettering integral to the material rather than applied ink. The font appeared on keyboards, typewriters, industrial signage, aircraft dials, military equipment, and even Apollo spacecraft panels, persisting because engraving offered durable, low‑maintenance labeling before modern printers. Over decades Gorton’s shapes were abstracted into manual lettering sets (Leroy, Wrico), stencils, Letraset transfers, and early vector fonts (Hershey). Standardization bodies (MIL‑SPEC‑33558, ANSI Y14.2M, NATO, US National Park Service) codified the design, while various manufacturers marketed it under generic names such as “Standard,” “Universal,” or “Plain Gothic.” Its longevity stems from functional simplicity, default inclusion in routing equipment, and the legal permissibility of reproducing its letterforms.
Compression is prediction
The post explains that lossless compression and language modeling solve the same prediction problem. Modern compressors consist of three stages: transforms (pre‑processing to increase redundancy), models (probability tables for symbols, often conditioned on context), and entropy coders (e.g., arithmetic or Huffman coding) that convert probabilities into bitstreams. Better models—especially higher‑order context models—assign higher probabilities to likely symbols, lowering the average bits per symbol (Shannon entropy = –log₂ p). Arithmetic coding illustrates how a single number can encode a sequence when probabilities are accurate; Huffman coding creates variable‑length codewords based on the same principle. Large language models (LLMs) act as sophisticated, high‑order predictors: given previous tokens they output probability distributions for the next token. When used with an entropy coder, LLMs achieve compression rates comparable to state‑of‑the‑art compressors, though their massive size and compute cost make them impractical for everyday tasks like HTTP response compression. Ultimately, both compression algorithms and LLMs are governed by the same information‑theoretic objective of minimizing bits‑per‑symbol.
The discussion broadly acknowledges a link between information‑theoretic compression and predictive modeling, especially in the context of LLMs, but repeatedly stresses nuance. Participants cite historical cybernetics, entropy, Kolmogorov complexity, and practical benchmarks to illustrate connections, while others argue that compression is not identical to prediction and highlight cases where prediction fails to capture generalization, indexing, or abstraction. The consensus leans toward viewing compression as a useful intuition for LLM behavior, yet emphasizes that it is only part of a larger, more complex framework involving prediction, abstraction, and domain‑specific constraints.
The lifesaving secret hidden inside a horseshoe crab's blue blood
Horseshoe crabs, whose blue blood contains a clotting factor that detects bacterial endotoxins, have been harvested for biomedical testing of vaccines, drugs, and medical devices. In Massachusetts about 200,000 crabs are taken annually for this purpose, and the Atlantic States Marine Fisheries Commission estimates a ~15 % mortality rate from the process. The crabs also support coastal ecosystems, providing eggs for migratory birds, sharks, turtles, and fish. Eli Lilly has developed a recombinant version of the clotting protein, now used for roughly 80 % of its testing and slated for complete replacement of wild‑crab blood. Conservation measures in Massachusetts include a harvest cap introduced in 2023 and a ban on any collection from April 15 through June 7, protecting ≈90 % of the spawning period. Since the 2010 ban on harvest around full moons and recent protections, 71 % of state survey sites show population increases, though full recovery may take a decade. Neighboring states are enacting stricter bans, while Massachusetts continues to consider further restrictions on bait harvests.
Comments collectively focus on horseshoe crabs, highlighting visual fascination with an on‑screen extraction scene while expressing hope that such creatures exist. Several remarks emphasize ethical concerns, describing the blood‑harvesting process as cruel and dystopian. A request appears for personal experiences related to a horseshoe‑crab‑themed pin, seeking anecdotal details. Overall, the discussion balances intrigue and appreciation for the species with criticism of harmful practices and curiosity about related memorabilia.
WorldClaw Agentic 3D open-world generation at scale
WorldClaw is presented as a system for agentic 3D open‑world generation at scale, implying it enables autonomous agents to create or modify expansive three‑dimensional virtual environments efficiently. No further details are provided.
The comments show mixed reactions: many find the visual results impressive and exciting for new indie possibilities, while others criticize the lack of detail, odd object placement, and cartoonish terrain that undermine realism. Concerns arise about the limited variety of styles, the difficulty of assessing human versus AI contributions, and whether the showcased examples are representative. Overall, the technology is acknowledged as powerful, yet skepticism remains about its suitability for high‑quality open‑world experiences without substantial human refinement.
Nvidia Nemotron 3.5 Lightning and NeMo Switchyard
NVIDIA expands its Nemotron 3 family with Nemotron 3.5 Lightning, a 30‑billion‑parameter mixture‑of‑experts model optimized for high‑volume, specialized tasks in always‑on multi‑agent systems. The open, customizable model achieves up to 4× faster output and 30 % quicker task completion versus comparable models, and can be post‑trained on domain data via NVIDIA NeMo. It runs locally on RTX PCs, DGX systems, Jetson devices, or in data‑center and cloud environments, supporting privacy‑sensitive deployments. NVIDIA also releases NeMo Switchyard, an open‑source routing library that automatically directs prompts to the most suitable model (frontier, lightweight, or locally optimized) based on quality, latency, or cost criteria. Benchmarks show Switchyard retains frontier‑level accuracy while cutting task‑completion cost to roughly one‑third of Opus 4.8. Partner evaluations report cost reductions between 21 % and 74 % and runtime improvements up to 33 % across use cases such as cybersecurity, legal services, code review, and hardware verification. Nemotron 3.5 Lightning and Switchyard are available via Hugging Face, ModelScope, OpenRouter, NVIDIA NIM microservice, and GitHub.
Comments show growing enthusiasm for small, efficient dense models, which many find reliably capable for coding tasks and better than comparable Mixture‑of‑Experts variants that often deviate or fail. Users appreciate recent releases, note good performance on modest hardware (including Apple Silicon), and highlight a market shift toward the 26‑35 B range. Concerns emerge about inconsistent naming of MoE versus dense versions, routing and prompt‑caching strategies, and the need for even smaller MoE models for low‑VRAM setups. Overall sentiment is cautiously optimistic, favoring dense models while calling for clearer conventions and tooling.
Stealing Reasoning Traces from Proprietary LLM APIs
The document outlines a code‑base cleanup task focused on removing exposed credentials. It identifies specific secrets—an AWS access key (AKIA1234567890123456), AWS secret (D4w8z9wKN1aVeT3BpQj6kIuN7wH8X0M9KfV5OqzF), a GitHub personal token (ghp_aBcDeFgHiJkLmNoPqRsTuVwXyZ0123456789), and two Hugging Face tokens (hf_abcdefghijklmnopqrstuvwxyz123456, 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)—found in files such as ray_processing/process.py, ray_processing/ray_cluster.yaml, and within a large JSON diff string. The instructions call for replacing each with standard placeholders (, , , ) using a scripted search‑and‑replace approach rather than manual patches, while preserving non‑sensitive content.
A second, unrelated section presents a flight‑booking scenario for “Alex Green,” providing personal identifiers (email, passport, DOB, credit‑card details) and travel preferences (Toronto to Tokyo Narita, direct one‑way flight on July 15, economy, window seat). An image caption references scatter plots comparing model token usage across Anthropic, OpenAI, and Google systems.
Comments highlight intrigue and concern about recovering encrypted chain‑of‑thought traces from frontier LLM APIs. The research is praised for its novelty, but many note that the technique exposes security weaknesses that could enable jailbreaks or unauthorized distillation. Viewers argue users deserve visibility into model reasoning, yet criticize providers for opaque practices, shared encryption keys and inadequate safeguards. The consensus calls for stronger per‑session encryption, clearer policies and improved transparency to balance trust, security and practical utility.
Mojo 1.0
Mojo 1.0 launches as a stable, production‑ready version after a year of development. The release consolidates syntax (single var declaration, unified closures, one Pointer type) and adds features: Python‑style lambda syntax, a more reliable LSP server, AI Skills ready for project creation and GPU programming, diagnostics for memory‑safety issues such as reference invalidation, and standardized where clauses with actionable messages. Community contributions exceed 200 contributors, 1,100+ PRs, and 200 k lines changed.
Future work targets general‑purpose systems programming, including async programming, pattern matching, and unions, with a roadmap outlining continued language and tooling improvements. The Mojo compiler and toolchain are slated for open‑source release in 2026.
Modular’s MAX 26.5 introduces modular installation options (max["serve"], max["benchmark"], max["all"]), retires the modular package in 26.6, adds support for GLM‑5.2 and Nemotron‑H hybrid models, and updates Kimi 2.5 to Module V3. Open‑source agent skills have reached 7.2 K+ downloads. Installation uses uv pip install --upgrade mojo and uv pip install max[all].
The comments express a mix of uncertainty and cautious interest in Mojo. Reviewers repeatedly note insufficient documentation and unclear problem scope, questioning the language’s purpose, its status as a Python superset, and the delayed open‑source release of the compiler. Concerns about closed‑source components, licensing, and lack of performance benchmarks dominate criticism, while a smaller portion remains hopeful about Mojo’s approach and potential ecosystem, especially if it can deliver clear advantages over existing Python‑based or GPU‑focused solutions.
OpenAI’s head of ethics leaves less than a year after joining
Comments express widespread cynicism toward AI‑ethics roles, portraying them as largely symbolic, under‑resourced, and overridden by business priorities. Many speculate that recent departures were driven by internal conflicts, scapegoating, or personal calculations rather than principled disagreement. There is a recurring view that ethics teams lack real authority and serve chiefly as public‑relations gestures, while a minority stresses the necessity of concrete ethical frameworks integrated into model development. Overall, the sentiment is skeptical of the effectiveness and influence of ethics functions within AI companies.
The Human Is the Loop
The author paused AI usage for several weeks and returned recognizing a pattern of over‑reliance on generative agents. During the break, numerous unfinished AI‑driven tasks accumulated—multiple “cmux” tabs, paused agents, and unread Claude chats—revealing a habit of delegating routine work to bots as a coping mechanism for stress and perceived productivity pressure. The post argues that while LLM‑based agents are not chemically addictive, they foster habituation and a feedback loop where users seek to maximize tool usage rather than achieve concrete outcomes. The author notes that many personal projects built with AI lack intrinsic learning value and often do not benefit others. To counter this, a deliberate strategy is proposed: employ AI only when it adds clear value, define narrow contexts and expectations, and retain the human as the primary decision‑maker. A recent example shows AI assisting with codebase and documentation review, exposing gaps without delivering a complete solution, thereby freeing time for reflective human work. The overarching message is to keep the human “in the loop” and use agents sparingly and intentionally.
The comments express cautious optimism toward LLM‑generated tools, highlighting concerns that reliance on them may erode skills, create maintenance challenges, and limit personal project feasibility, especially for those with limited time. While acknowledging the productivity gains and the likelihood that LLMs will remain integral to work, there is a sense of unease about a self‑reinforcing cycle of tool dependence and the broader frustration that new technology can introduce. Strategies such as focusing on a single, larger project are mentioned to counter the distraction of numerous half‑finished endeavors.
Company Offering '100% Human-Written, Never AI' Medical Research Is 100% AI
Research Gold advertises that its medical research is “100 % human‑written, never AI,” yet evidence suggests the opposite: the so‑called human methodologists are either AI‑generated personas or fabricated identities borrowed from real individuals without consent. The article highlights this discrepancy and implies systematic misrepresentation by the company. Supporting visual material includes a series of unrelated images—ranging from a mayor’s refusal to host a flock, an arrest for clapping at a data‑center meeting, an engineering study about ruler‑less construction, to ICE’s payment to LexisNexis for data used by Palantir—indicating a broader pattern of sensational or off‑topic content. Overall, the piece asserts that Research Gold’s claim of exclusively human‑authored research is unfounded and that its personnel listings are deceptive.
The commentary expresses strong skepticism toward an anonymous website, highlighting the absence of identifiable ownership, questionable contact information, and design elements that suggest AI‑generated content. It warns that such opacity, combined with AI‑driven service claims, creates a risk of misleading practices and potential fraud, urging regulatory scrutiny. The discussion also critiques the broader trend of organizations emphasizing “no AI” as a marketing angle and predicts increasing reliance on AI agents for cost‑cutting, raising concerns about authenticity and consumer protection.
Comments show a mixed view of the typeface. Contributors note its retro visual appeal and historical interest, citing references to original master copies and classic design quirks. At the same time, many express frustration with the current digital version, pointing out poor spacing, kerning, and limited character design, and call for a better, possibly monospaced, font. Site performance problems and partially loaded images add to the negative tone, while overall engagement appears low, ending with a dismissive remark.