Unsloth Dynamic 3.0 GGUFs: +10% Accuracy on Qwen3.8-27B at Same Size

Unsloth launches Dynamic v3.0 for Qwen3.8-27B: +10% top-1% accuracy vs competitors, Divergence-300 @32 and KL Divergence improved, no QAT/QAD — per unsloth

DeepMind unveils WeatherNext: an extra day of warning for cyclones

DeepMind's AI model predicts cyclones with an extra day of accuracy and is now open source.

Magic hexagons of all orders: AI discovery

Researcher uses AI to find magic hexagons of all orders up to 21, breaking previous record.

Oberon System migrates from RISC-5 to RISC-V in new project

Project migrates Wirth's Oberon System to RISC-V, using OP2 compiler and RV32 emulator, staying faithful to the original book.

Domain for sale? Now DNS alerts with _for-sale record

New RFC 10023 standard allows announcing domain sale via DNS without taking down the site. Learn how it works.

NASA Extends Voyager 2's Life by One More Year with Power Adjustment

48-year-old probe gets a reprieve with energy changes; scientific instruments remain active.

Nixpkgs core team dissolves and criticizes Steering Committee governance

Nixpkgs central team announces dissolution, cites micro-management and poor SC communication; governance future open.

The Existential Crisis of Knowledge Workers in the Age of AI

Knowledge workers question the meaning of work in the face of AI. An analysis of the end of workism.

Bioengineered chewing gum reduces HPV by up to 93% in tests

Bioengineered chewing gum reduces HPV by up to 93% and eliminates bacteria linked to head and neck cancer, preserving the oral microbiota.

Kitesurf: Browser for AI Agents Runs in V8 Isolates on Cloudflare

Cloudflare launches Kitesurf, a browser for AI agents in Workers, with superior efficiency to Chromium.

US loses 23,000 jobs in July; labor market stagnates

US jobs report shows loss of 23,000 positions in July, with wages below inflation and labor force participation low.

Anthropic pays 6x market for events while CEO criticizes financial motivation

Dario Amodei fears new hires only care about money, but job pays up to $400k.

Nepal is the 47th Government to Join the Free Have I Been Pwned Service

Nepal's NCSC gains access to monitor governmental domains against data breaches on the HIBP platform.

New Mexico Court Orders Meta to Pay $567M for Child Harm

New Mexico court adds $567M to Meta's fine, totaling $942M for children's mental health harm.

AMD acquires Taalas and etches models into silicon to boost inference

AMD acquires startup Taalas to accelerate AI inference with model-specific circuits, generating up to 17,000 tokens per second.

Inouye reveals Kelvin-Helmholtz instability on the Sun's surface

Inouye telescope confirms vortices in the solar photosphere, paving the way to understand solar flares.

Kimi K3: open model with 2.8 trillion parameters arrives in July

Kimi K3, from Moonshot AI, is the first open source model in the 3 trillion parameter class, with 1M context and native vision.

OpenAI Releases 100,000 Access Grants to Advanced Models for Scientists

OpenAI grants free access to 100,000 academic researchers to accelerate science and collaboration.

How to Create a Nintendo 64 Game in 2026: The Saga of Xibalba 64

Developer ported JavaScript engine to C and created an FPS for N64, with physical release by Modretro.

Muse Code and Spark 1.2: Meta's terminal agent arrives in beta

Meta launches Muse Code, a terminal coding agent, powered by the Muse Spark 1.2 model, focused on complex tasks and kernel optimization.

Pareto in Mario Kart: how mathematics defines the best kart

Antoine Mayerowitz uses Mario Kart 8 to explain the Pareto frontier and multi-objective optimization.

The 4 Simple Rules Behind Japan's Most Livable Cities

Japanese cities lead in quality of life with accessible design and efficient transit. Understand the rules.

Entropy of Markov chains: from physics to artificial life

How Boltzmann inspires a definition of entropy for Markov chains, applied to Dyson's cell model.

Intelligence is not the main bottleneck, says essay on medicine and AI

Essay argues that regulation and clinical trials, not intelligence, hold back medical advances even with AI.

Quantego: IBM quantum computers become LEGO models

Quantego project recreates IBM Quantum System One and System Two in LEGO, with instructions and quantum circuit simulation.

Cloudflare OS: open platform for agents and apps at work

Cloudflare open-sources Cloudflare OS, a platform that gives every person an agent and workspace with company context.

Qwen 3.0 Image Pro: image generation with text in 12 languages

Alibaba's new model generates images with legible text in 12 languages, supports up to 4.5k tokens, and costs from US$ 0.003.

Lean kernel bug refutes Collatz: lessons in soundness

Ramana Kumar announced a verified refutation of the Collatz conjecture in Lean, but the proof had a kernel bug. Lessons on soundness and proof objects.

Pi proves that minimalism in code harness cuts costs and maintains quality

Databricks study and Shopify case show that the minimalist Pi harness outperforms competitors with lower cost.

Shieldstral: 3B multimodal moderator that outperforms models 7x larger

Mistral releases Shieldstral, a 3B multimodal safety classifier under Apache 2.0, outperforming models up to 7x larger.

The Golden Age of British Ice Cream: 1976-1991

How 15 years transformed British culture: from Cornetto to Magnum, the industrial revolution of ice cream.

ADLC Team Skills: team patterns for Claude Code and Codex

Open-source project brings engineering standards to AI agents like Claude Code and Codex, with versioned guidelines.

Stateless MCP 2.0 reignites interest and inspires new tools

New stateless MCP specification simplifies implementation and leads to creation of mcp-explorer and datasette-mcp.

DeepSeek V4 Flash runs on a single AMD MI300X with 168 tok/s

Repository shows how to run DeepSeek V4 Flash on an AMD MI300X, with 168 tok/s and 256K context, without quantization.

Germany sets record: 12 billion kWh of solar power fed into the grid in July

Germany records a historic high of 12 billion kWh of solar energy fed into the grid in July 2026, according to IWR.

Technical expertise is the key to getting the most out of LLMs

Experts achieve superior results with LLMs. Understand why domain knowledge is the new critical skill.

The House That Didn't Know Humanity Died: August 4, 2026

Bradbury's tale of an autonomous house in 2026 gets a new reading: the date has arrived and fiction becomes a warning.

Andy Pavlo leaves CMU to lead the new ClickHouse Labs

CMU database professor Andy Pavlo joins ClickHouse to found ClickHouse Labs and drive database research.

Don't Be an AI Middleman: The Value Is in Understanding

Passing on AI responses without understanding is useless. Learn to add value by reading, validating, and writing in your own words.

MiniMax H3 arrives in ComfyUI: open weights, native audio, and 2K video

MiniMax H3, an open-weights video model, now runs in ComfyUI with stereo audio and 2K output.

Why Hollywood Stopped Filming in Hollywood

Analysis of the migration of film production from Los Angeles and its impact on film quality.

BitNet on a BBC Micro: language model in 13KB of weights

Project runs an autoregressive BitNet model on a 1980 BBC Micro, with 13KB of weights and inference code in C.

CP/M-386: Classic CP/M Reborn in 32-bit Protected Mode

Project brings CP/M to 386 protected mode with Ring-3 TPA, boot from floppy or GRUB, and high compatibility with CP/M-68K and 2.2.

Locksmith Loop Agent Migrates COBOL to Java with 91.9% Coverage

Agentic method validates COBOL-to-Java migration with deterministic oracle, achieving 91.9% coverage in a real case.

SSH becomes a collaborative drawing canvas: ssh ssh.place

ssh.place turns SSH into a collaborative canvas: 200x60 cells, no account, one pixel every 15s.

Bor v0.8.0: open source policy management for Linux with new types

Bor v0.8.0 adds Thunderbird, Edge and Firewalld, revamps UI and strengthens security.

F*: the proof language that protects Azure and Firefox

Meet F*, the open source language from Microsoft Research and Inria used in verified cryptography in Azure, Firefox, and Linux.

Portable Fasttracker II clone in C with SDL 2 reaches v2.22

Developer releases clone of Fasttracker II in C with SDL 2, version 2.22, open source on GitHub and support for Linux, Windows and macOS.

RISC OS Open celebrates 20 years of open source and eyes 64-bit

ROOL completes two decades of opening RISC OS; Apache 2.0, Raspberry Pi, and Moonshots plan mark the journey.

Go 1.27 arrives with generic methods, UUID, and post-quantum signature

Go 1.27 brings generic methods, UUID, ML-DSA, allocation optimizations, and leaked goroutine profiling.

Kimi K3 on MI355X: AMD Beats B300 in Performance per Dollar

AMD MI355X runs Kimi K3 at 952 tok/s per node, beating B300 in performance per dollar. Wafer.ai details optimizations.

RipGrep Musl Binaries Crash on Large Searches

Segfault incidents reported during extensive searches with RipGrep 15.2.0 on x86_64-unknown-linux-musl.

Seedance 2.5: 30s videos in one pass and multimodal references

ByteDance launches Seedance 2.5 with 30s generation, multi-round extension, and multimodal references for professional creation.

A Deeper Understanding with 'The Art of 64-bit Assembly'

Understand the inner workings of programs with the book 'The Art of 64-bit Assembly.'

BMW Introduces Film Advertisements on Car Displays

On July 27, 2026, BMW began displaying Spider-Man advertisements on car displays.

Liberal Arts Education and Its Applications in Career and Life

The skills acquired in liberal arts courses have a surprising impact on technical careers, such as programming.

The Renaissance of Personal Computing with the Help of Artificial Intelligence

Discover how AI is revolutionizing personal computing, enabling the creation of highly personalized applications without the need to scale.

$500 Fine-tuning Surpasses Frontier Models in Catalog Review

With RL on an open 9B model, the company surpasses GPT-5.5 and Claude Opus 4.8 in legal task at a fraction of the cost.

Defcon 34 badge brings open source chip that becomes a security key

Baochip-1x chip, created by bunnie Huang, is open source and can be inspected down to the silicon.

Google fixed more Chrome bugs in June than in two years, with AI

With AI agents, Google fixed 1072 Chrome security bugs in two months, surpassing 23 previous versions.

Zig and GTK4: a lightweight ssh-askpass for Gentoo without X11

Developer creates ssh-askpass in Zig and GTK4 to avoid X11 and KDE dependencies on Gentoo.

Andrew Ng launches LearnVector with $100M from Coursera for one-on-one tutoring with AI

LearnVector, Andrew Ng's new company, receives $100 million from Coursera to create one-on-one learning experiences with AI, with products expected by earl

Anthropic CEO denies advocating ban on open-weights models

Dario Amodei clarifies position: does not support banning open models, but advocates chip controls and safety testing.

Using an open model is surprisingly liberating, dev reports

Modal engineer tells how running Kimi K3 on his own endpoint brought a sense of freedom and control over data.

Harmony Explained: Towards a Scientific Theory of Music

2012 article proposes a computational theory of harmony, deriving the major scale, standard chord dictionary, and the difference between major and minor tr

Kimi Linear Outperforms Full Attention with 75% Less KV Cache

Moonshot AI's hybrid linear attention architecture outperforms full attention in all scenarios, with efficiency gains and open source.

Lean 4 proves CSG correction with 93 lines of specification, ignoring AI code

verified-3d-mesh-intersection project uses Lean 4 to verify 3D mesh intersection. 93-line specification eliminates review of 1000+ lines of AI code.

RTX 2080 Ti gains 22 GB of VRAM in memory upgrade

Service doubles RTX 2080 Ti VRAM capacity to 22 GB, ideal for AI and heavy workloads.

Misago abandons React.js and adopts HTMX to simplify forum

Open source forum project Misago removes React.js from codebase and adopts HTMX, eliminating template duplication and improving performance.

Neutrino-1 8B: 3.88 GB ternary model runs on GPU and laptop

Fermion Research launches Neutrino-1 8B with ternary weights that fit in 3.88 GB, serving GPU, Mac and CPU from the same artifact.

France faces unprecedented 'pyrocumulonimbus' in giant fire

Fire in southwestern France generates fire cloud with lightning and its own winds, a phenomenon never seen in the country.

Lean + LLMs: Automated Proofs Make Type Systems Viable

Google engineer shows that LLMs can automate proofs in Lean, drastically reducing the effort of formal verification.

Applications of AI in Legal Education: The University of Chicago Approach

The University of Chicago Law School adapts its legal education to the AI era, creating policies and guidelines for AI use.

Decker: multimedia platform rescues HyperCard legacy with open source

Decker is an open source multimedia platform that rescues the spirit of HyperCard, with Lil language and ditherpunk aesthetic.

GrapheneOS details defenses against data extraction from locked devices

GrapheneOS explains its protection layers: rate limiting, insider attack resistance, duress PIN, and automatic reboot for data security.

Ruff v0.16.0: 413 default rules, Markdown formatting, and new suppressors

Ruff v0.16.0 raises default rules from 59 to 413, formats Python blocks in Markdown, and adds suppression comments.

Inflect-Micro-v2: Advancing Speech Synthesis Under 10M Parameters

Inflect-Micro-v2 offers full English speech synthesis with a fixed model, capable of handling long texts and running on CPU or CUDA.

The Mighty Semicolon in Shell Scripting: Efficiency and Pragmatism

Discover how using the null command and parameter expansion in the shell can improve your efficiency.

Updated Context Engineering Rules for Claude 5

Claude Code has evolved: discover how to update your context engineering with more advanced generation models.

W4ME Station Enables Running WASM-4 Games on Java ME Phones

W4ME Station, a new WASM-4 runtime for Java ME devices, brings unmodified WebAssembly games to CLDC 1.1/MIDP 2.0 devices.

98.css: CSS library recreates Windows 98 interfaces on the web

98.css is a CSS library that faithfully recreates the appearance of Windows 98. No JavaScript, compatible with any framework, and focused on accessibility.

ascdraw: native and open-source ASCII/UTF-8 diagram editor reaches 120+ FPS

Free GPLv3 tool for text diagrams, with infinite canvas, layers, and exportation in TXT/JSON/PNG formats.

Claude Opus 5 Leads Artificial Intelligence Analysis Index

Claude Opus 5 achieves the highest score in the Artificial Analysis Intelligence Index, surpassing 170 evaluated models.

Developing a Miniature 3D Renderer for the Playdate

See how the challenge of creating a 3D software renderer for the Playdate handheld device went.

Fedora 45: A Study on Package Manufacturing

An inside look at the creation process of Fedora 45 artifacts, from the package commit to the final installable version.

OpenAI launches GPT-5.6 Sun with a 54% increase in token efficiency

Sam Altman reveals that the new model is 54% more efficient in agentic coding tasks and as good as rivals.

Postgres LISTEN/NOTIFY: Efficiency and Scale in Connectivity

How Postgres LISTEN/NOTIFY achieved 60K writes per second with millisecond latency.

Prof. Hannah Fry Receives the Leelavati Prize for Mathematics Popularization in 2026

Hannah Fry honored at the International Mathematicians Conference for mathematics popularization work.

Researchers reveal the secret of vertical sleep in sperm whales

New research from the University of St Andrews and the Université de Neuchâtel discovers how sperm whales sleep vertically

Sam Altman Reveals AI 2026: Predictions for OpenAI and Tech Industry

AI 2026 roadmap unveils predictions for OpenAI's future, including advancements in GPT-6 and shifts in consumer and enterprise AI demands.

Security in security cameras questioned after GitHub token leak

GitHub admin token discovered in Hanwha security camera firmware, exposing hundreds of repositories.

Spatial Languages: Writing 2D Code Opens Up New Dimensions in Programming

Explores how the addition of an extra dimension can revolutionize the way we write and understand code.

Textual fingerprint analysis reveals similarity between Kimi and Claude

Typebulb tool compares models by writing patterns; Kimi (Moonshot) and Claude (Anthropic) appear as the most similar.

Wasmtime 47 Releases Support for GC and Exceptions in WebAssembly

Wasmtime version 47 brings GC and exceptions by default, expanding the reach of WebAssembly to more programming languages.

GigaToken promises tokenization 1000x faster than HuggingFace

New Rust library achieves GB/s tokenization, outperforming HuggingFace and tiktoken in benchmarks.

SIMD Simplified: Why Every Developer Should Know This Technique

Mitchell Hashimoto shows that SIMD isn't just for experts: with 5 simple steps, any dev can speed up loops by up to 16x.

Apollo 11 Guidance Computer Source Code Available on GitHub

GitHub repository contains the original AGC code for the command and lunar modules, digitized by the MIT Museum.

Kimi K3 surpasses GPT-5.6 and Opus 4.8 in agentic benchmark, but costs a lot

Chinese 2.8T parameter model achieves second highest Elo on AA-Briefcase, behind only Claude Fable 5, with a cost of US$10.57 per task.

NEO Radar: Engine Tracks 41,812 Real Asteroids Precisely

New orbital tool simulates all planets with high fidelity, using precise algorithms for tracking space objects.

Qwen vs DeepSeek vs GLM: Guide to the Three Chinese AI Giants (2026)

Technical comparison between Qwen3.6, DeepSeek V4 and GLM-5.2: benchmarks, licenses, prices and usage recommendations.

2026 Guide: which open-weight model to choose for each task

GLM 5.2, DeepSeek V4, MiniMax M3, Kimi K2.6 and Qwen 3.6 compared in intelligence, cost, context and licensing.

Altman proposes 5% of OpenAI for US sovereign wealth fund while launching GPT-5.6

Sam Altman launches GPT-5.6 with aggressive pricing and suggests giving 5% of OpenAI equity to American citizens via a sovereign wealth fund.

China's Kimi K3: The AI Sputnik Moment Has Arrived

Emad Mostaque and Peter Diamandis analyze the launch of the Chinese K3 model, marking a turning point in the global AI race.

Kimi K3 shakes markets and reignites AI price war

Chinese model with 2.8 trillion parameters rivals GPT-5.6 and Opus 4.8, sending AI and semiconductor stocks down.

Moonshot AI's Kimi K3: 2.8T Open-Source Model Takes Lead in Front-End Coding

Moonshot AI launches Kimi K3, a 2.8 trillion parameter open-source model, surpassing Claude Fable 5 on Frontend Code Arena with a 1M token context window.

Open-weight LLMs 2026: DeepSeek, Qwen, Kimi, GLM and Llama Head-to-Head

Technical comparison of price, context, license and performance of the leading open models in 2026, with emphasis on Chinese ones.

PCjs: IBM PC Emulator in Browsers Revives the 70s and 80s

PCjs brings classic hardware emulation via JavaScript, including IBM PC, minicomputers, and arcades. Open source project on GitHub.

Fast inference changes what you can build - Insights

Emad Mostaque: The Plan to Save Humanity From AI

Emad Mostaque outlines his vision for a decentralized, sovereign artificial intelligence infrastructure ('Intelligent Internet'), the concept of Universal Basic AI (UBAI), and technical solutions to prevent economic collapse.

Emerging Situation: Anthropic's Global Pause, Recursive Self-Improvement, and AI Personhood Arrives

A deep dive into Anthropic's 'When AI Builds Itself' report, the technical reality of systems coding their own evolution, and Argentina's pioneering framework for AI personhood.

Geoffrey Hinton on AI Consciousness, Oracle Safeguards, and the 'Fog of Exponential Progress'

AI pioneer Geoffrey Hinton shares new perspectives on machine sentience, alternative safety models, and why predicting the impact of AI beyond a decade is like driving in a heavy fog.

DeepSeek V3 Outperforms GPT-4o on Coding and Math — At a Fraction of the Cost

China's DeepSeek V3 has surpassed OpenAI's GPT-4o on several key benchmarks while costing 10x less per million tokens — a seismic shift in the AI cost curve.

Kimi K2: Moonshot AI's Bet on Agentic Models Challenges OpenAI's Operator

Moonshot AI's Kimi K2 introduces a new class of long-context, agentic models capable of multi-step reasoning and tool use — positioning China as a leader in autonomous AI agents.

Alibaba's Qwen3 235B: The Largest Open-Source Model — And It's Chinese

Alibaba's Qwen3 235B-A22B is the largest open-source language model available, outperforming GPT-4o on multiple benchmarks while being freely downloadable and self-hostable.

The Chip War in 2026: How China Is Routing Around US Semiconductor Sanctions

Despite sweeping US export controls on advanced semiconductors, China's AI industry has accelerated — not slowed. Here's how Beijing's strategy is working.

Andrej Karpathy: Software 3.0, Agentic Engineering, and Why You Can't Outsource Understanding

At AI Ascent 2026, former OpenAI co-founder Andrej Karpathy lays out the transition to 'Software 3.0', the rise of vibe coding, and the critical distinction between building and understanding in an AI-first world.