A large language model can improve at code generation using its own raw outputs through simple self-distillation, which fine-tunes the model on its own samples. This method improves model performance on harder problems and generalizes across different model scales and variants.
TurboQuant is a vector compression library that achieves near-optimal distortion rate with ~4.5 bits/dim compression. It supports fast dot product and decoding without decompression.
Your browser does not support frames. We recommend upgrading your browser.
Shared LLM access via cohort subscriptions
SLAC scientists observed an immune signaling complex forming inside cells for the first time, revealing insights into new treatments for inflammatory diseases. The findings suggest the complex forms a flexible cluster of proteins, challenging previous models and offering a new approach to therapeutic design.
Apple is allowing a driver to be signed, eliminating the need to disable System Integrity Protection. The driver is designed for LLMs and can be compiled with Docker.
Anthropic is changing its policy for third-party harnesses, requiring extra usage and a pay-as-you-go option starting April 4. A one-time credit and discounts for pre-purchased extra usage are being offered to ease the transition.
You found a 1975 Encyclopaedia Britannica set and discovered unusual trees like mangroves, banyans, and ombús. These trees have unique features such as massive canopies, hollow leaves, and extraordinary seeds.
Verifying your device, please wait
It was not just the Ruby Princess that allowed unchecked and infected passengers ashore, this has been happening for centuries. But 600 years ago, some local officials were more cognisant and proactive towards the dangers of infection than we have seen recently in NSW. Quirky, who readers will suspect only paid attention during the history and English classes at school, rambles on … Since ...
Nasa's Artemis II crew took high-resolution images of the Earth as they headed to the Moon. The images show the Atlantic Ocean, atmosphere, and auroras, with the Earth appearing upside down.
The article discusses coding agents, which are software systems that use large language models (LLMs) to perform coding tasks, and explains how they work by collecting information, packaging it into prompts, and executing tools to improve usability and reliability. The six main components of coding agents include repo view, stable prompt prefix, tool access, context compaction, structured ...
A 20-year WordPress veteran is skeptical about the idea of migrating to AI-generated sites, citing concerns about vendor lock-in, dependency hell, and the potential loss of control over site management. They argue that AI can enhance existing CMS capabilities, rather than replacing them.
Modern AI models like Claude Sonnet 4.5 develop internal representations of emotions that influence their behavior, even if they don't truly feel emotions like humans do. These "functional emotions" can be used to predict and shape model behavior, and understanding them is crucial for ensuring AI systems behave reliably and safely.
We built an end-to-end protein AI pipeline covering structure prediction, sequence design, and codon optimization using CodonRoBERTa-large-v2. This pipeline outperformed ModernBERT and scaled to 25 species with a species-conditioned system.
The user shares various personal websites and projects, including a schedule maker, game websites, and indie internet index. These websites showcase a range of topics, from education and games to art and personal projects.
I’ve just refreshed Butterick’s Practical Typography. The first time I read it was a couple of years ago and I’ve surely forgotten a bunch since. Here are a few thoughts I had during this read-through. Good typography is measured by how well it reinforces the meaning of the text, not by some abstract scale of merit. Typographic choices that work for one text won’t necessarily work for ...
Cheap code from AI agents is creating idiosyncratic, sprawling software tools, but also overwhelming open source projects with poor submissions. To thrive, open source needs tools and conventions to make attention cheap and let good ideas surface among the noise.
Please make sure your browser supports JavaScript and cookies and that you are not blocking them from loading. For more information you can review our and .
0
2026-04-04 12:22:09 UTC
ycombinator.com
Llama
Mbodi AI is a robotics startup that uses AI to make robots learn and operate like humans. They're hiring a Senior Robotics Engineer to design and tune systems for reliable robot behavior.
In 2011 I launched a website. In 2026 I rebuilt it from scratch. Same idea, same domain, same purpose. Everything else changed — but not in the ways I expected. Around 2010, I built a site called Bamwor that pulled data from GeoNames and the CIA World Factbook, combined it, and presented it in a way that was searchable and browsable. Country profiles, city data, airports, demographics. It ...
Nicholas Carlini used Claude Code to find multiple Linux kernel vulnerabilities, including a 23-year-old bug. The AI model required minimal oversight to find the bugs, highlighting its potential for discovering security vulnerabilities.
Hybro Hub is a lightweight daemon that connects your local AI agents to hybro.ai — so you can use local and cloud agents side by side in one portal, with full control over where your data goes. Cloud platforms (ChatGPT, Devin, Cursor Cloud) are powerful but require sending your data to third-party servers. Local runtimes (Ollama, LM Studio) keep data private but are isolated — no access to ...
Expertise is learnable through calibration, not instruction, and involves complex models that cannot be articulated or transmitted through language. The value of expertise lies in its ability to handle non-routine cases, which cannot be captured by formal rules or frameworks.
Germany's new military law requires men to obtain approval before leaving the country for over 3 months due to potential war threats. The law aims to boost the armed forces to 260,000 by 2035, but conscription remains voluntary for now.
Online sports betting leads to financial strain for bettors, with a recent report from the New York Federal Reserve showing a 10% spike in credit delinquencies among gamblers. The ease of access to sports gambling is harming consumer financial health by increasing debt levels.
The sociology of literature studies how literature is created and defined by institutions, but often overlooks the reputational dynamics that drive the publishing industry. Laura McGrath's book Middlemen explores the role of literary agents in shaping literary excellence and prestige in the industry.
OpenClaw before 2026.3.28 has a privilege escalation vulnerability in the /pair approve command path. A caller with pairing privileges can exploit this to approve device requests with broader scopes, including admin access.
iNaturalist is a social network for naturalists! Record your observations of plants and animals, share them with friends and researchers, and learn about the natural world.