LLMs reward expertise

https://www.seangoedecke.com/og-image.jpg
Using large language models like LLMs requires domain knowledge to get the most out of them. A skilled prompter like Terence Tao can steer the model with expertise in the domain, but this skill is not easily replicable by following tips.

Show HN: Run an 80B Qwen in 4.3 GB of RAM on a Mac, and a 35B on an iPhone

https://opengraph.githubassets.com/33c2323be10c8fc55b1ca1698918168f8c99940b20885c0e596079d869780dac/leonickson1/Swiftlet
Swiftlet is a Swift + Metal runtime for Qwen3-Next and Qwen3.5/3.6 models, allowing them to run on iPhone with 2.5 GB of RAM. It streams weights from storage and activates only 3B parameters per token.

Amazonian civilization had estimated 3M people in 3% of forest area

Ten advances in mathematics and theoretical computer science

Apple is getting this wrong

Devtools must be open source

https://blog.exe.dev/static/og-card.png
Software engineers used to rarely write personal software, but now agents can easily manage customizing software, making it easier to personalize and maintain. Agents can build and manage custom software with minimal programming, allowing for more efficient and flexible use of software.

CollectWise (YC F24) Is Hiring

https://bookface-images.s3.us-west-2.amazonaws.com/logos/a5be02a3378f664e1f80f2a84997514d00f9fc1f.png?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=ASIAQC4NIECAGJ4GR23X%2F20260804%2Fus-west-2%2Fs3%2Faws4_request&X-Amz-Date=20260804T070953Z&X-Amz-Expires=3600&X-Amz-Security-Token=IQoJb3JpZ2luX2VjED8aCXVzLXdlc3QtMiJHMEUCIQD1ZJP59c5kdnjtfy%2BPJb1R5bTO1MpH0JTjR4302ocetQIgFQyNSOVWMvboON2AGPrzKAzFHKKpuUAyJMDaVDvSApkq5QMICBAAGgwwMDYyMDE4MTEwNzIiDNxIQsPyoKcC9dfhFirCA4QpZid%2BnNu%2FWATxtevBeAUZ9%2FQgfioggfUc%2FZYcF7psnla9OBOgh0uUIgx5L3KIbNc26sysSCNkkg03oOS8y7h7nw9aCS31jVkmdL%2BSEiE7U0TNFR58Xnn2jFlM4NN7kbEcGzmyY2AnMFOH9xdvFAufQbd%2BSgIytJvdgElzzCrgBSD3Ws%2FAltf3kdvj2906mJ3YfLV2T0NJlSrdKIhG%2BNvYNkBe467Pcy%2FvUGKmxXM4kvNTi5IEK0k6rGjKjQndxi2n1Mb2F14%2FCeI0k463bqZpG4LoBQ459Fk8nvSXtI69G9nrKIhlIusy1LQLzHPDlZmDO1KrG3NuU3L0VRIG4HFYX3NcmVM8%2Bh2YbCjpcVx74BLdbyhxYSwdZl2wyRilEbczdwUPu8due6JkOi7Kox%2Fg0LYSy%2Bj3v3yvqxJJflPBbMCL2ycESK03C3QR7fPGIbv3%2B0%2FovNZiguo4MPl65SY2exFz4Bi7S5MTDevYOGZ%2B2mgQfgijlFjHjkOZh3lVvuVlhbq3VnqynY2MMGnt8q1XCJblhk9oT%2BD%2FoAwhhXbQxxsjqbmX0XTVFClAgYtYEhwuEW%2BZ9HN9Dky6VjFzZLnxpTDhmMbTBjqlAYGb8ycelRH0%2B%2FWdJ9vpPqljuFFZAnTzSsvlJs1%2FRfk3i9YJUAqaWz9ZI5ibLG3GG6W7U7eJJLS0KmRlJjng7pyAkhBfIjo2pfOS9W1CqBEVqEee9OlvB56q9YL9neEz4Nb7eLHT3NP%2FC5rTcuLLAKp4FNvQRh%2BMvNipmlGPXdjBmlkfNiBW74mMeIpEWIQeQkwQuYYLY55gXjZ1sYQN2Q9yPKbHLw%3D%3D&X-Amz-SignedHeaders=host&X-Amz-Signature=ad4f832ad80c404fa11faed841717b557d4b3294867617d348d6f8f3a3c818ef
CollectWise is a Y Combinator-backed startup using generative AI to automate debt collection. They're hiring an AI Agent Engineer to build voice AI infrastructure and prompting systems to reach $10 million in annual revenue within the next year.

There Will Come Soft Rains (1950) [pdf]

Ask HN: Who is hiring? (August 2026)

Bucket Robotics: Senior Fullstack Engineer, San Francisco, CA (On-site) - Own massive surface area, build customer-facing app, scale internal tooling, and design data pipelines. Product-minded engineer with systems thinking and curiosity.

That time when I failed the Microsoft interview

https://ochagavia.nl/images/headshot.jpg
The author, a Computer Science student, applied for a Microsoft internship in 2015 after contributing to open source projects. They received a phone interview invitation, prepared using a book on Big Tech interviewing practices, but displayed honesty when unable to solve a brain teaser, which may have affected their chances.

Smaller, faster, safer: running Kimi and GLM at scale

https://blog.cloudflare.com/_image?href=https%3A%2F%2Fblog.cloudflare.com%2F_emdash%2Fapi%2Fmedia%2Ffile%2F01KZ1PVA5SC9J87PS1R2B97N1F.png&w=1999&h=1125&f=webp&fit=cover&position=center
Workers AI optimizes large models on GPUs by quantizing the KV cache to FP8, compressing model weights to INT4, and protecting the cache with integrity checks. These techniques enable efficient serving of models like Kimi K-series and GLM with no change in accuracy, supporting more customers at lower costs.

Ask HN: Who wants to be hired? (August 2026)

Several professionals with AI, data, and software engineering backgrounds are seeking roles in various locations, including Brazil, Canada, India, Russia, and the US, with expertise in technologies like Python, SQL, Generative AI, LLMs, and Rust. They are looking for opportunities as Senior AI Engineers, Applied AI, Search/Relevance, RAG, Data/AI Lead, and AI Solutions roles, with some open ...

Prevent cognitive debt by manually retyping LLM-generated code

The user uses coding assistants to fast-forward through boring parts of projects, but manually types generated code to understand and adapt it, valuing comprehension over productivity. This workflow helps them build a mental model of their codebase and detect potential issues, allowing them to work faster and more efficiently.

Beauty in my backyard

https://worksinprogress.co/.netlify/images?url=https%3A%2F%2Fassets.worksinprogress.co%2Fwp-content%2Fuploads%2F2026%2F07%2F471416621_122181002012057398_8649519096196337300_n.jpg&w=1080&h=1080&q=65&fit=cover
Western countries used to let landowners build almost anything anywhere. Now, strict rules freeze the character of most neighborhoods, banning change and densification. Why? One theory is that modern buildings tend to be ugly, leading residents to support rules that exclude them from their neighborhoods. If this is true, then designing beautiful buildings could lead to more development. ...

MiniMax H3 Day-0 Support in ComfyUI: Open Weights, Native Audio, and 2K Video

https://substackcdn.com/image/fetch/$s_!59Bc!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F98d2552b-d84c-4385-9602-bf344dca2ebe_1000x727.png
MiniMax H3 is a next-generation open-weights video model that generates 2K video with real stereo sound up to 15 seconds. It combines text, images, video, and audio inputs with a prompt to create a single video clip, reducing five tasks into one model.

Launch HN: Hoplite (YC S26) – Effortlessly deploy cloud coding agents

https://hoplite.sh/opengraph.png
Effortless cloud coding agents that feel good to use.

Andy Pavlo joins ClickHouse to establish ClickHouse Labs

https://clickhouse.com/_next/image?url=%2Fuploads%2Fandy_pavlo_clickhouse_b978ca187a.png&w=750&q=75
I'm joining ClickHouse to lead ClickHouse Labs, a research team focused on advancing database technology. Our goal is to conduct impactful research and transform ideas into technology that benefits users, building on ClickHouse's strong foundation and exploring emerging AI and agentic technologies.

Twenty Years of Pandoc

John MacFarlane created the document converter pandoc in 2006 using the Haskell programming language, which he chose for its strong type system and clean representation of structured documents. Over the next 20 years, pandoc evolved into a powerful tool supporting 51 input and 76 output formats, with over 600 contributors and a reputation for reliability and determinism, despite the potential ...

Celebrating 45 Years of Kermit with the First New C-Kermit Release in 15 Years

https://changelog.complete.org/.within.website/x/cmd/anubis/static/img/pensive.webp?cacheBuster=1.22.0
Please wait a moment while we ensure the security of your connection.

IPC-7351B Electronic Component Zero Orientation [pdf] (2010)

The industry needs a consistent technique for electronic component orientation and land pattern geometries to facilitate data capture and transfer. A fixed Zero Component Orientation is proposed to achieve reliable solder joint formation platforms and improve efficiency in electronic product development automation.

Windows XP 2002 for the Itanium: Unbridled rage

200 Milliseconds

https://200ms.thenodebook.com/og.png
See what really happens in the ~200ms after you press Enter: a single HTTP request traced in real time through DNS, TCP, TLS, the Linux kernel, Node's event loop, and Postgres — then back to the pixel. Scroll distance equals time.

Replacing the Kobo Libra H2O Battery

https://ei3lh.eu/wp-content/uploads/2025/11/6b6b3f7e-9b96-4505-ab02-f2d7569cb651-1024x558.jpg
User had a Kobo Libra H2O with a faulty battery, which was difficult to replace due to lack of suitable options. They found a replacement battery on Ali Express and successfully replaced it with a soldering iron and basic tools.

Why did we wait so long for the bicycle? (2019)

https://blog.rootsofprogress.org/img/giovanni-fontana-carriage.png
The invention of the modern bicycle required decades of design iteration and technological advancements, including improved metalworking, wheel technology, and road quality, to make it practical and safe. The development of the bicycle was also influenced by general economic and cultural factors, such as the need for a middle class to provide demand and the availability of surplus resources ...

How Hollywood stopped making movies in Hollywood

https://substackcdn.com/image/fetch/$s_!PU5R!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ed3afc1-5576-4677-aa1a-bf941416e9d7_1548x1024.webp
The entertainment industry in Los Angeles has been experiencing a production migration due to rising costs and tax incentives offered by other cities, leading to a decline in physical production in Hollywood. This migration has resulted in a widening gap between where stories are set and where they are filmed, with many productions choosing locations based on financial considerations rather ...

Bonsai: Janestreet's UI Library

https://raw.githubusercontent.com/janestreet/bonsai/master/docs/assets/bonsai-logo.png
Bonsai is a UI library for building performant, reactive web applications in OCaml, inspired by Elm. It allows for composable state machines and incremental rendering, making it easy to manage state and UI updates.

Decades-old fish sauce at abandoned factory in Canada finally being removed

https://lede-admin.defector.com/wp-content/uploads/sites/28/2024/12/ugly-baby-salad.jpg?resize=2880%2C2880
A Canadian town, St. Mary's, has been plagued by a 20-year-old fish sauce stench from abandoned vats, but cleanup efforts have finally begun. The resulting fish sauce, if processed, might have a unique flavor profile due to prolonged fermentation, with notes of rancidity, amines, and possibly sweetness.

ZX Spectrum System Tour: Text Mode

https://bumbershootsoft.wordpress.com/wp-content/uploads/2026/05/spectrum_manic_mechanic.png?w=625
The author is exploring the ZX Spectrum's machine language capabilities, discussing how to load and run machine code, interact with the system variables, and print text to the screen. The author also covers how to read keyboard and joystick input, and provides a simple example program that demonstrates these capabilities, paving the way for more complex projects like the Manic Mechanic game.

More German than many Germans

https://cdn.synaps.media/mertbulan/content/images/size/w1920/2026/08/IMG_8938.webp
The author, a Turkish immigrant, had a positive experience living in Germany, finding a friendly flatmate and a supportive work environment. He eventually became a German citizen after a 14-month process, feeling a sense of belonging and contributing to the society.

AirLLM 70B inference with single 4GB GPU

https://raw.githubusercontent.com/lyogavin/airllm/main/assets/airllm_logo_sm.png
AirLLM is a library that reduces inference memory usage for large language models, allowing them to run on a single GPU card without quantization or pruning. It supports various models, including Llama, Qwen, DeepSeek, and more, and can be used with a simple one-line initialization.