I extracted the safety filters from Apple Intelligence models

https://opengraph.githubassets.com/c7bacf0afff793b7f5733a7e82a390aad095fd66304a509a53bd391959481a3f/BlueFalconHD/apple_generative_model_safety_decrypted
To run the decryption script, install cryptography using pip and attach Xcode's LLDB to GenerativeExperiencesSafetyInferenceProvider.

LLMs should not replace therapists

https://arxiv.org/static/browse/0.3.4/images/icons/smileybones-pixel.png
Large language models are not suitable to replace therapists due to their inability to form a therapeutic alliance and express stigma towards mental health conditions. LLMs lack human characteristics necessary for a therapeutic relationship, making them unsuitable for replacing human therapists.

Building the Rust Compiler with GCC

https://fractalfir.github.io/images/gcc_bootstrap_process.png
The user is working on bootstrapping the Rust compiler using GCC instead of LLVM, which involves fixing several issues including incorrect implementation of 128-bit SwitchInt terminator and misaligned pointer dereference. The user was able to identify and fix these issues, but encountered a new problem with vectorization passes causing undefined behavior, and is still investigating the cause.

Show HN: I wrote a "web OS" based on the Apple Lisa's UI, with 1-bit graphics

https://alpha.lisagui.com/media/social.png
A Lisa-themed Graphical User Interface
The author is annoyed by the increasing number of AI-related Show HN posts on Hacker News, which they believe are low-effort and self-promotional. They used SQL queries to analyze the data and found that AI-related posts have increased significantly, with a notable spike in 2023.

Why English doesn't use accents

https://substackcdn.com/image/fetch/$s_!uLxv!,w_520,h_272,c_fill,f_auto,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F252d7601-59bd-4f35-a045-a5ce13df8e20_1564x1269.jpeg
The Norman Conquest of 1066 introduced French influence on English, leading to the use of extra letters to spell sounds and the eventual loss of diacritics. English retained this system, while France developed diacritics to distinguish between sounds, creating a historical irony in their differing writing practices.

Jane Street barred from Indian markets as regulator freezes $566 million

https://image.cnbcfm.com/api/v1/image/108167923-17516114451751611442-40559065982-1080pnbcnews.jpg?v=1751611444&w=600&h=300&vtcrop=y
SEBI temporarily barred Jane Street Group from India's securities market, accusing it of widespread market manipulation. SEBI froze 48.4 billion Indian rupees from Jane Street and directed banks to prevent any transactions without permission.

A non-anthropomorphized view of LLMs

In many discussions where questions of "alignment" or "AI safety" crop up, I am baffled by seriously intelligent people imbuing almost magical human-like powers to something that - in my mind - is just MatMul with interspersed nonlinearities.In one of these discussions, somebody correctly called me out on the simplistic nature of this argument - "a brain is just some ...

Async Queue – One of my favorite programming interview questions

https://davidgomes.com/content/images/2025/07/CleanShot-2025-07-06-at-14.10.27.png
The user describes a programming interview that tests a candidate's ability to implement an async queue with a sendOnce function that guarantees only one request is handled by the server at a time. The interview also tests the candidate's ability to handle new requirements, write elegant code, and work with AI tools, making it a challenging and interesting assessment of a software engineer's skills.

Intel's Lion Cove P-Core and Gaming Workloads

https://substackcdn.com/image/fetch/$s_!sSwz!,w_520,h_272,c_fill,f_auto,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F699136d2-f7b3-425b-a65d-ae52c83d9070_1384x1600.jpeg
Lion Cove is Intel’s latest high performance CPU architecture. Compared to its predecessor, Raptor Cove, Intel’s newest core can sustain more instructions per cycle, reorganizes the execution engine, and adds an extra level to the data cache hierarchy. The list of changes goes on, with tweaks to just about every part of the core pipeline. Lion Cove does well in the standard SPEC CPU2017 ...

Get the location of the ISS using DNS

https://shkspr.mobi/blog/wp-content/uploads/2025/07/1024px-International_Space_Station_17_April_2002.jpg
The user discussed using DNS LOC records to store server locations, including latitude, longitude, and altitude. They created a proof-of-concept to update these records every 15 minutes using an API.

Functions Are Vectors (2023)

https://thenumb.at/Functions-are-Vectors/fadd.svg
The article discusses the concept of representing functions as infinite-dimensional vectors, allowing the application of linear algebra tools to various problems. It explains how functions can be viewed as vectors with infinitely many values, and how this perspective enables the use of linear algebra techniques such as matrix multiplication and diagonalization. The article also explores the ...

Backlog.md – CLI that auto-generates task files (took my Claude success to 95 %)

https://raw.githubusercontent.com/MrLesk/Backlog.md/main/.github/backlog.gif
Backlog.md turns any Git repo into a project board using Markdown files and a zero-config CLI. It provides a modern web interface for task management with human-readable data saved as Markdown files.

Opencode: AI coding agent, built for the terminal

https://raw.githubusercontent.com/sst/opencode/dev/packages/web/src/assets/lander/screenshot.png
To install opencode-ai, use a command like curl -fsSL https://opencode.ai/install | bash or install via package managers like npm or brew.

Nobody has a personality anymore: we are products with labels

https://substackcdn.com/image/fetch/$s_!fur_!,w_520,h_272,c_fill,f_auto,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F32309b8c-9b23-43db-b70a-534c68ce1dcc_1262x980.png
Therapy-speak has taken over our language. It is ruining how we talk about romance and relationships, narrowing how we think about hurt and suffering, and now, we are losing the words for who we are. Nobody has a personality anymore. In a therapeutic culture, every personality trait becomes a problem to be solved. Anything too human—every habit, every eccentricity, every feeling too ...

I don't think AGI is right around the corner

https://substackcdn.com/image/fetch/$s_!cAOq!,w_520,h_272,c_fill,f_auto,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2a0dc7f9-1224-47b3-9508-d56bfd9fe14f_1029x562.jpeg
The user is bearish on transformative AI in the next few years due to the lack of continual learning in current LLMs, but bullish on AI over the next decades when this capability is solved. They predict a broadly deployed intelligence explosion around 2032, with AI capable of learning on the job and handling complex tasks like computer use and white collar work.

Lessons from creating my first text adventure

https://xkqr.org/if/lockout/lockout.jpg
The author created a text adventure game called Lockout, which took 15 hours to make and 57 hours to fix, and learned that making text adventures is harder than it seems due to technical challenges and the need for a detailed world model. The author also discussed the importance of scope, game design, and parser development in creating a fun and immersive text adventure experience.

Metriport (YC S22) is hiring engineers to improve healthcare data exchange

https://bookface-images.s3.amazonaws.com/small_logos/d4da680d3cdee4cc7ad6408d3177026092f04c75.png
Metriport is a healthcare data intelligence platform that integrates with major healthcare IT systems, providing real-time access to patient data. They're hiring full-stack engineers with entrepreneurial spirit and a strong work ethic to join their team in San Francisco.

Corrected UTF-8 (2022)

Corrected UTF-8 fixes design errors in UTF-8 by removing overlength encodings, skipping C1 controls and Unicode surrogates, and removing the artificial upper limit on code space. It adds offsets to codepoints to ensure unique encodings.

Cool People [pdf]

A second factor discussed within the literature is autonomy. being rough, rebellious, ironic, hedonistic, autonomous, calm, adventurousness emerged as traits that are more cool than desirable. identifying these attributes would give us utmost insight as to why cultures value coolness, have embraced the concept, and why it has spread with its meaning intact. cool (vs. not cool) people were ...

Hannah Cairo: 17-year-old teen refutes a math conjecture proposed 40 years ago

https://imagenes.elpais.com/resizer/v2/5LK4ENJC3ZBI3EFXNAIFAIDHJY.jpg?auth=8911f2a3f3ba941e6ca304ceaac3c5c7cdb155253ee7ed064b6b7e6459525312&width=414
Hannah Cairo, a 17-year-old, solved the Mizohata-Takeuchi conjecture, a decades-old math problem, using fractals and careful arrangement. She was a high school student when she solved the problem, which was widely believed to be true, and was greeted with surprise and enthusiasm by the math community.

Mirage: First AI-Native UGC Game Engine Powered by Real-Time World Model

https://blog.dynamicslab.ai/diagrams/mirage_training_diagram.png
Mirage is a real-time generative engine using AI World Models for live UGC gameplay. It enables dynamic, interactive, and sustained gameplay through natural language, keyboard, or controller input.

Toys/Lag: Jerk Monitor

You created a tool to measure latency in computer interactions, specifically for wireless mouse dongle delays and browser rendering. It shows frame and pointer movement delays, as well as event batching and offsets.

Collatz's Ant and Σ(n)

https://gbragafibra.github.io/gifs/norma_score3.png
The score function Σ(n) is normalized by the stopping time τn, resulting in a ratio. The ant's distance to the origin is measured by α and γ, with a ratio of γ/α.

Overclocking LLM Reasoning: Monitoring and Controlling LLM Thinking Path Lengths

https://royeisen.github.io/assets/figures/new_teaser.jpg
Researchers studied how large reasoning models track thinking progress and found that hidden states encode a token's relative position. They developed a method to extract and modify this representation, allowing for dynamic reasoning control and real-time interpretability.

1945 TV Console Showed Two Programs at Once

https://spectrum.ieee.org/media-library/photo-of-a-wooden-tv-console-with-two-cathode-ray-tube-television-sets.jpg?id=61133345&width=1200&height=1004
Allen B. DuMont invented dual-screen TVs in the 1940s and 1950s to allow multiple people to watch different programs. His vision of separate yet together viewing was ahead of its time, but never went into mainstream production due to economic reasons.

Serving 200M requests per day with a CGI-bin

https://static.simonwillison.net/static/2025/cgi-card.png
Jake Gold tests 90s-era CGI with a Go + SQLite program on a 16-thread AMD 3700X, serving 2400+ requests per second. CGI works well today with modern languages like Go and Rust, taking advantage of multiple CPU cores.

Hidden interface controls that affect usability

https://interactions.acm.org/images/uploads/article_images/3735949/ins01.gif
Philip Kortum argues that modern devices often require users to have knowledge in their head to perform tasks, which is a step backward from the earlier design principle of knowledge in the world. He advocates for designers to create more usable systems with visible controls that can be accessed without prior knowledge, citing examples of successful designs in mission-critical systems and ...

Can we test it? Yes, was can [video]

Show HN: Simple wrapper for Chrome's built-in local LLM (Gemini Nano)

https://opengraph.githubassets.com/f5ff639dcf82e126193d1e086a84f93dcdbf2a36941b67096f1440fe0433a3df/kstonekuan/simple-chromium-ai
The simple-chromium-ai library provides a lightweight TypeScript wrapper for Chrome's AI Prompt API, offering simplicity and type safety. It handles initialization, session management, and error handling for AI model usage.