1. The tiny corp raised $5.1M
Total comment counts : 46
Summary
The author has started a new company that aims to commoditize “the petaflop” by building an open-source framework, runtime, and driver for AMD chips. The company’s short-term goal is to get AMD on MLPerf using their tinygrad framework. They believe that starting an AI chip company requires starting with the software and that the computing in ML is not general-purpose computing. The article outlines the advantages of their tinygrad IR, which has 12 operations, all of which are ADD/MUL only. The author expresses a desire for an open-source framework that can run cutting edge AI on any chip as seamlessly as Linux supports 50 network cards.
Top 1 Comment Summary
George Hotz, a well-known hacker, has announced a new project called “PicoCUDA” as his latest technology endeavor. Hotz has previously created knockoff versions of popular technology, gained media fame, but then failed to deliver due to a lack of resources or focus. However, PicoCUDA aims to be an alternative to CUDA/NVIDIA, a widely-used parallel computing platform and API, which could potentially break Hotz’s pattern of fizzled-out projects. Despite this, individuals are advised to be cautious before giving Hotz a deposit.
Top 2 Comment Summary
The author suggests that any AI chip company should start with software, since most models used in machine learning have their compute and memory accesses statically computable. He also praises Nvidia’s focus on software, including CUDA, Tensor cores, and complementary software like TensorRT. Finally, the author expresses a desire for more players in the space.
2. Everything you always wanted to know about mathematics (2013) [pdf]
Total comment counts : 28
Summary
error
Top 1 Comment Summary
The author is asking for advice on the best way to read PDFs because their ebook reader doesn’t support it, and they don’t like reading long-form content on a computer screen. One suggestion given by a commenter is to use a service to print and bind PDFs into a physical book. The author is seeking other suggestions to improve their reading workflow.
Top 2 Comment Summary
The article discusses a book that was used for a first-semester course for computer science and math undergrads at Carnegie Mellon University in the past. The course currently uses a different resource listed at a provided URL.
3. SectorC: A C Compiler in 512 bytes
Total comment counts : 21
Summary
SectorC is a C compiler written in x86-16 assembly that fits within the 512 byte boot sector of an x86 machine. It is capable of supporting a subset of C that is large enough to write real and interesting programs. The article describes the author’s journey in developing SectorC, including overcoming the challenges of fitting a C tokenizer/lexer within the limited memory space and developing a Barely C Programming Language that minimizes the number of tokens. Ultimately, the author was able to create a proper C compiler with spare memory to include additional features. The article also discusses the lack of error checking in traditional C programming and provides examples of programs leveraging the unique hardware aspects of the x86-16 IBM PC.
Top 1 Comment Summary
The article discusses a C-subset compiler and JIT interpreter that lacks operator precedence, but is still an impressive technical achievement. The author suggests a small optimization to improve code density in 16-bit code.
Top 2 Comment Summary
The article mentions the concept of compilers bootstrapping, where a slightly more advanced version of a programming language can be compiled with a simpler compiler, and then with each subsequent iteration, the language becomes more powerful, ultimately leading to advanced compilers like GCC. It also references the SectorC language, which can be compiled with the TCC compiler.
4. Ask HN: How do you not take criticism of your work personally?
Total comment counts : 336
Summary
The article discusses the importance of developing an anti-fragile attitude towards criticism to become a better developer and human being. This involves increasing awareness of one’s emotional state and not tying self-worth to output or delivery. The writer argues that self-worth should be tied to how well individuals live their values. They also suggest attaching self-worth to doing good work for a cause rather than the work itself.
Top 1 Comment Summary
The article explains that one of the necessary steps to become a senior developer is to develop an anti-fragile attitude towards criticism and handle it in a way that helps you become a better developer. The article compares fragility, non-fragility, and anti-fragility, and suggests that you need to expose yourself to smaller shocks to develop anti-fragility. The author also suggests that increasing your awareness of your emotional state can help you handle criticism and become a better human being overall.
Top 2 Comment Summary
The article suggests a few tips for improving communication, especially in asynchronous textual communication. The first tip is to assume positive intentions from the other person, defaulting ambiguity to the more pleasant side and adding kindness to the conversation. The second tip is to not respond immediately, giving yourself emotional distance and time to see and fix the problem. The article suggests focusing on feeling better and not on making the other person feel worse. Lastly, it suggests reflecting on your reactions to calm yourself down and realize the situation’s unimportance.
5. How to Finetune GPT-Like Large Language Models on a Custom Dataset
Total comment counts : 14
Summary
error
Top 1 Comment Summary
The author asks whether it is possible to use a tool like GPT-4 to label data and use that data as a training set for their own LLM. However, the author notes that the OpenAI Restriction Terms of Service state that users cannot use output from the service to develop competing models.
Top 2 Comment Summary
The article reports that training the Vicuna on a custom dataset could be easy, according to a link provided that discusses how to train the Vicuna with Skypilot.
6. Show HN: Image background removal without annoying subscriptions
Total comment counts : 46
Summary
Pixian.AI is a background removal service that enables customers to remove backgrounds from images. It uses powerful GPUs and multi-core CPUs to analyze images and remove the background from them. The service is free while in Beta and is expected to come out of beta in Q2 or Q3 of 2023. The company intends to offer a free tier for low-volume users, and pay-as-you-go style credit packs, with no monthly fees or minimums, and no subscriptions when they come out of beta. Pixian.AI’s pricing model is designed to maximize quality and minimize price while enabling graphics professionals, hobbyists, and ad-funded apps and websites to use the background removal service at a ‘makes sense’ price.
Top 1 Comment Summary
The article provides information on two addons available on Github that can be used with auto1111, which is a software that allows users to blur and redact sensitive information. One addon generates masks using Facebook’s SAM, while the other addon does the same task well. The article explains how to quickly install these addons by pasting their Github URLs into the “install from URL” box.
Top 2 Comment Summary
The article states that records not associated with an account are deleted or anonymized within a year of creation while image processing records associated with an account are retained indefinitely. The reason behind retaining the records is to allow users to view and download them anytime they wish and to provide customer support. However, the author doesn’t seem interested in retaining the records and politely declines the offer.
7. VanJS (Vanilla JavaScript): smallest reactive UI framework
Total comment counts : 44
Summary
VanJS is a reactive UI framework based on pure vanilla JavaScript and DOM that helps manage state and UI binding with a simple API, offering every good aspect that React does without the need of React, JSX, transpiling, or any hidden logic. The framework is very thin and delegates most of the work to standard browser APIs implemented in native code, resulting in a small bundled size of just 1.2 kB. VanJS provides first-class support for TypeScript and puts heavy emphasis on simplicity, with only 4 exported functions in the API. The goal of VanJS is to reduce the entry barrier for UI programming and build a better world, with no intention or plan for commercialization.
Top 1 Comment Summary
The article argues that using “VanJS” as an abbreviation for “Vanilla JavaScript” could cause confusion as the term already exists to refer to JavaScript without any framework. The author suggests using alternative puns such adding a scoop of vanilla to your project or sprinkling in some Reactive syntax instead.
Top 2 Comment Summary
The article describes a DSL for manipulating the document object model (DOM) on webpages, and a system that replaces the DOM from scratch each time data changes. This approach is criticized for being bad for performance and accessibility, as it involves recomputing style and requiring repaints even for minor changes. The author suggests that the supposed benefit of smaller bundle sizes does not outweigh the drawbacks of abandoning rendering engine computations.
8. Tape Storage Trundles On, Increases Yearly Volume to 128 Exabytes
Total comment counts : 24
Summary
Despite frequently being declared “dead”, tape storage has seen a prolonged increase in shipments recently, with a 5% YoY increase from the Linear Tape-Open Program Group which includes IBM, Quantum Corporation and HPE. The latest technology, LTO-9, has greatly increased compressed data density up to 45TB and increased transfer rates of up to 900 MB/s. Shipping volumes of tape storage, in general, have seen a decline but recent increases suggest that as a long-term storage solution, tape still has many benefits.
Top 1 Comment Summary
The author’s dissertation research from 1980 on Jupiter’s upper atmosphere is stored on a 2400 foot reel of tape in their attic. The data is approximately 40 megabytes, but the author has no plans to try to rebuild the scattering matrix for Jovian cloud particles.
Top 2 Comment Summary
The use of tape for backup has been declared as dying by those who do not use or understand the differences between tape and disk. However, the assertion that drives do not fail often enough to justify the use of tape is incorrect. In order to be safe, data should be on at least two drives, which increases the cost. The use of older LTO drives and tapes is both easy and cost-effective for backing up data.
9. AI Canon
Total comment counts : 36
Summary
error
Top 1 Comment Summary
The article discusses articles on the topic of generative AI posted by a16z, which can be found by clicking on the domain of the submission. It also provides a link to a page from a couple of years ago with submissions related to DAOs.
Top 2 Comment Summary
The article provides a list of essential resources for general AI studies, including Norvig & Russel’s textbook, Goodfellow’s Deep Learning, Koller & Friedman’s PGMs, the Krizhevsky ImageNet paper, the original GAN, and the AlphaGo paper and Atari DQN paper.
10. Archive of medieval books and manuscripts discovered in Romanian church
Total comment counts : 9
Summary
A team of researchers from Babeș-Bolyai University have discovered over 200 books and manuscripts dating as early as the ninth century in a church tower in Mediaș, central Romania. Among them are 139 printed books from between 1470 and 1600, two manuscripts from the 16th century and around 60 charters, records and documents between the 14th and 16th centuries. Furthermore, there were manuscripts retained inside parish records, the oldest of which was from the Caroliginian period and may date to the ninth century. St Margaret’s Church, or Margarethenkirche, was founded by Transylvanian Saxons in the 15th century and the collection seems to have been left inside the church’s tower for at least decades, possibly for protection during WW1 or WW2. The team will now match up the books with those listed in a catalogue from 1864, which contained around 7,700 volumes.
Top 1 Comment Summary
The article discusses the excitement of discovering forgotten or hoarded texts through modern imaging techniques. These texts offer insight into trade links from monastic scribe houses and may contain valuable DNA fragments, as well as possibly strange pictures of cats and snails.
Top 2 Comment Summary
The article discusses the need to digitize medieval books and manuscripts so that scholars can research the findings collectively. However, the article wonders about the cost of this digitization process and which research labs can provide this service.