Hasty Briefs - Daily Tech News Brief

Bilingual

You Can Now Disable All AI Features in Zed
9 months ago
- Zed aims to be the world's best code editor, with first-class AI features.
- Users can now disable Zed AI via a global setting in settings.json.
- New users will soon have an option to disable all AI features during onboarding.
- Reasons to disable AI include concerns about training data, environmental impact, and machine-generated code.
- Organizations may restrict AI tool usage for legal or proprietary code reasons.
- Privacy alternatives include using your own API keys or local AI models.
- Zed AI service ensures code and prompts are not stored or used for training.
- Zed acknowledges skepticism around AI but encourages understanding its role in software development.
- Zed is open-source (GPL license), allowing further customization.
- Future goals include Windows support and improving both AI and non-AI experiences.
The Last 100k Hacker News Links
9 months ago
- Egg intake may be associated with Alzheimer’s dementia risk in older adults.
- Immutable Linux distributions enhance security and stability by making core file systems unchangeable.
- Retro computing nostalgia sparked by Commodore revival announcement.
- QSBS Planning Calculator helps estimate tax savings for startup equity gains.
- FastStripe integrates with Answer.AI for an improved Stripe experience.
- GitHub Spark enables software development with over 150 million users.
- Analysis of Myanmar earthquake reveals curved fault slip captured by CCTV.
- AI coding assistants tested for migrating NgRx to signals, with mixed results.
- Problem Spotter uses AI to discover product opportunities in Reddit communities.
- NVIDIAScape vulnerability shows container misconfigurations can compromise GPU hosts.
- CT-ScanGaze dataset and baselines for 3D volumetric scanpath modeling.
- LookAway 1.14 syncs with iPhone to block distractions during breaks.
- FreeBSD 15.0 installer to include KDE Plasma desktop environment option.
- Portmaster v2 enhances privacy and security by controlling network connections.
- Obamas discuss productive framing for the male malaise crisis.
- Auto-scaling limitations in serverless architectures and alternative solutions.
- Carbon and price-aware Kubernetes scheduler optimizes deployment costs and emissions.
- Robot 'cannibal' grows stronger by consuming smaller robots.
- Guide to using Qwen3-Coder with Claude Code via LiteLLM.
- Airlines increasingly charge solo travelers more.
- President Trump’s AI Strategy focuses on innovation and education.
- WebAssembly boosts HTML5 game performance.
- Q2 net outages caused by governments, power cuts, and cable damage.
- Tech industry’s AI integration criticized for degrading software quality.
- Google’s AI tool Aeneas predicts missing words in Roman inscriptions.
- Fifth potentially habitable planet found around red dwarf L 98-59.
- Cloudflare’s Jetflow framework improves data pipeline efficiency by 100x.
- Valid HTML zip bomb explained.
- Comparison of Xian and Vyper for Web3 smart contracts.
- Intel Simics 6 transitions to legacy state.
- Trump’s AI Action Plan outlines innovation and infrastructure recommendations.
- UK accidentally invaded Spain in 2002, revealed by former First Sea Lord.
- Sam Altman discusses AI and OpenAI on Theo Von’s podcast.
- Female moths use plant acoustic emissions in oviposition decisions.
- Zams automates CRM updates and sales tasks for B2B teams.
- Trump eases regulations on coal and chemical industries.
- Helix 25.07 text editor new features.
- The Promised LAN advocates for building local networks.
- Apple Health MCP server queries health data with natural language.
- The Elements of Artificial Intelligence series by Brady Gerber.
- Suzuki FRONX global automobile lineup.
- Interactive programming in C explored.
- Science behind coincidences, including lottery wins.
- Zero-knowledge proof article on sgo.to.
- Apple’s FastVLM improves vision language model efficiency.
- Seth Godin’s thoughts on seeing the lottery.
- Personal reflection on 'I Have vs. I Am'.
- Cyata explores AI roles in engineering workflows.
- UN court allows countries to sue over climate change.
- Comparison of file, object, and block storage solutions.
- ProdE AI acts as a production engineer for coding teams.
- IRS loses 25% of IT staff since Trump took office.
- Coca-Cola’s sugar push benefits sugar industry.
- Kimi K2 vs. Claude 4 Sonnet for agentic coding.
- UCLA improves radiant cooling for outdoor comfort.
- Imideo turns images into AI-powered videos.
- Apache Fluss is a streaming storage for real-time analytics.
- Base58 vs. Base85 encoding differences explained.
- Floor796 features a huge animation scene with memes and references.
- Delegation-oriented FedCM for web identity.
- Debian’s security update processes detailed.
- AI may not recursively self-improve, part 2 analysis.
- Brainbase Labs introduces Kafka as first AI employee.
- Hacker plants wiping commands in Amazon’s AI coding agent.
- Neuralagent AI lives on desktop and uses it like a user.
- BuildTogether reimagines collaborative coding with language models.
- Space Force tests OCX software blocks for GPS satellites.
- Converge team hiring in NYC.
- Ramparts scans MCP servers for vulnerabilities.
- Zed editor adds option to disable AI features.
- Brain Box text adventure adaptation.
- Ministry of Testing discusses tool bundling trends.
- Non-invasive neuromotor interface enables high-bandwidth computer input.
- The universe has no center, expanding in all directions.
- Thinking Machines’ $2B funding signals next AI wave.
- VMware portal issues prevent some users from downloading patches.
- AI friend apps criticized for societal impact.
- FT article on new Coca-Cola version requires subscription.
- Building Hoot reflection.
- Matrix.org recovers from Postgres corruption.
- Benchmarking RAG systems playbook.
- Symbol.so offers copy-paste text symbols.
- AnkiTTS leverages Eleven Labs API for text-to-speech.
- Microsoft’s Semantic Telemetry optimizes LLM interactions.
- Serverless single tenant RAG with DuckDB.
- Exponential improvement for Ramsey lower bounds.
- Toad universal UI for agentic coding in terminal.
- WebDAV viewer handles large text files efficiently.
- Volonaut Airbike raw flight footage.
- Embedded LEDs in silicone soft robots.
- Chess.com bans 100,000 cheaters monthly.
- Viral black plastic kitchen utensils paper corrected again.
- Quantum levitation physics explained.
- Homeless ER visits decrease after moving into King County hotels.
- Whisper app converts voice thoughts to text.
- Saucy automates GitHub contributor payouts.
- SQLite AI extension for on-device inference.
- Millet failed to become staple in ancient Japanese kitchens.
Human teens beat AI at an international math competition
9 months ago
- AI models achieved gold-level scores at the International Mathematics Olympiad (IMO) for the first time.
- Google and OpenAI's AI programs solved 5 out of 6 IMO problems, earning gold medals.
- Complex mathematics remains a significant challenge for AI due to its reliance on token prediction rather than logical reasoning.
- AI's performance at the IMO showed major improvements, with models explaining their solutions clearly.
- Despite progress, AI still falls short of the perfect scores achieved by some human competitors.
- Concerns persist about the massive energy and resource consumption required for AI advancements.
How big tech is force-feeding us AI
9 months ago
- Big tech is aggressively pushing AI products onto users through various design and interface strategies.
- A new study highlights how companies like Google, Meta, and Adobe use deceptive design patterns to impose AI features.
- AI features are often given prominent placement in interfaces, interrupting regular workflows and encouraging interaction.
- Examples include Google AI Overview, SnapChat's MyAI chatbot, and Adobe Photoshop's generative AI prompts.
- The study calls this 'forced use' of AI, where adoption is driven by companies rather than consumer demand.
- Silicon Valley's narrative of organic AI demand contrasts with the reality of manufactured ubiquity.
- OpenAI's nonprofit published a report with lofty, vague proclamations about AI's future, drawing criticism.
- The report's style and content have been mocked for being overly aspirational and lacking substance.
White House unveils plan to "win" global AI race through deregulation
9 months ago
- The White House released 'Winning the Race: America's AI Action Plan,' outlining the Trump administration's AI strategy.
- The plan aims for US dominance in AI through deregulation, infrastructure investment, and international partnerships.
- Critics argue the plan favors Big Tech by removing regulatory barriers and reversing Biden-era AI safety policies.
- AI is described as a catalyst for industrial, information, and cultural revolutions, likened to the 1960s space race.
- The plan's three pillars are innovation, infrastructure, and international diplomacy, focusing on accelerating AI adoption.
- It directs federal agencies to revise regulations hindering AI and reviews FTC investigations to avoid burdening AI innovation.
You can now train a 70B language model at home
9 months ago
- Answer.AI releases a fully open-source system to train 70b large language models on desktop computers with gaming GPUs (RTX 3090 or 4090).
- The system combines FSDP (Fully Sharded Data Parallel) and QLoRA (Quantized Low-Rank Adaptation) to enable efficient training on consumer hardware.
- QLoRA allows training by quantizing model weights to 4 bits and using LoRA adapters, reducing memory usage while maintaining performance.
- FSDP enables parallel training across multiple GPUs by sharding the model, avoiding the inefficiency of sequential processing.
- The project aims to democratize AI by making large model training accessible without expensive data center hardware.
- Key collaborators include Tim Dettmers, Hugging Face, and Answer.AI, leveraging open-source tools like bitsandbytes, PEFT, and Transformers.
- The system supports techniques like gradient checkpointing, CPU offloading, and Flash Attention 2 to optimize memory and performance.
- HQQ (Half-Quadratic Quantization) is introduced as an alternative to bitsandbytes, offering faster and more accurate quantization.
- Practical steps for using FSDP/QLoRA are provided, including installation and running training scripts on multi-GPU setups.
- The project is a first step toward more accessible AI model training, with future improvements and community contributions expected.
Anthropic teams use Claude Code
9 months ago
- Anthropic teams use Claude Code to transform workflows, automate tasks, and bridge skill gaps across various departments.
- Data Infrastructure team uses Claude Code for Kubernetes debugging, plain text workflows for finance, and onboarding new hires with codebase navigation.
- Product Development team leverages Claude Code for fast prototyping, synchronous coding for core features, and building Vim mode autonomously.
- Security Engineering team applies Claude Code for complex debugging, Terraform code reviews, and documentation synthesis.
- Inference team relies on Claude Code for codebase comprehension, unit test generation, and machine learning concept explanations.
- Data Science and ML Engineering teams build sophisticated dashboards and handle repetitive refactoring tasks with Claude Code.
- Product Engineering team uses Claude Code for workflow planning, independent debugging, and model iteration testing.
- Growth Marketing team automates ad creative generation and builds Figma plugins for mass production using Claude Code.
- Product Design team implements front-end changes directly and uses Claude Code for rapid prototyping and edge case discovery.
- RL Engineering team employs Claude Code for feature development, test generation, and Kubernetes operations guidance.
- Legal team builds custom accessibility solutions and automates legal workflows with Claude Code.
Qwen3-235B-A22B-Thinking-2507
9 months ago
- Qwen3-235B-A22B-Thinking-2507 introduces significant improvements in reasoning tasks, including logical reasoning, mathematics, science, coding, and academic benchmarks.
- The model features enhanced general capabilities like instruction following, tool usage, text generation, and alignment with human preferences.
- It supports a 256K long-context understanding capability, making it ideal for complex reasoning tasks.
- Model specifications include 235B total parameters, 22B activated, 94 layers, and 128 experts with 8 activated.
- Performance benchmarks show state-of-the-art results in knowledge, reasoning, coding, alignment, agent tasks, and multilingualism.
- The model supports deployment via Hugging Face transformers, sglang, vLLM, and other frameworks like Ollama and LMStudio.
- Best practices recommend specific sampling parameters and adequate output lengths for optimal performance.
- Qwen3 excels in agentic use with tool-calling capabilities, facilitated by Qwen-Agent for reduced coding complexity.
- Citations are encouraged for those who find the work helpful, with a reference to the Qwen3 Technical Report.
Google Opal
9 months ago
- Google Labs introduces Opal, an experimental tool for building AI mini apps using natural language and visual editing.
- Opal simplifies and visualizes workflows by chaining prompts, AI models, and tools without requiring coding knowledge.
- Users can create, tweak, and share AI apps, with options to edit via visual editor or conversational commands.
- Launched in a US-only public beta, Opal includes a demo gallery with starter templates for easy onboarding.
- Opal aims to empower creators to transform prompts into functional mini apps for prototyping, productivity, and more.
Efrit: A native elisp coding agent running in Emacs
9 months ago
- Efrit is an AI coding assistant integrated with Emacs, offering multiple interfaces for different tasks.
- Interfaces include efrit-chat for multi-turn conversations, efrit-do for natural language commands, and efrit-agent-run for automation.
- Features include direct Elisp evaluation, multi-turn conversations, tool integration, and safety-first design.
- Requires Emacs 28.1+, an Anthropic API key, and internet connection.
- Installation involves cloning the repository, adding to Emacs config, and setting up the API key.
- Usage examples include creating buffers with poetry, modifying content, and executing multi-step tasks.
- Customizable settings include model selection, token limits, and debug options.
- Troubleshooting tips cover common issues like API key errors and connection timeouts.
- Development includes build, test, and installation procedures, with contributions guided by CONTRIBUTING.md.
- Latest version (v0.2.0) includes stability improvements, enhanced token limits, and better documentation.
Nvidia Launches Family of Open Reasoning AI Models: OpenReasoning Nemotron
9 months ago
- NVIDIA announced the open Llama Nemotron family of models with enhanced reasoning capabilities for AI agents.
- The models improve accuracy by up to 20% and inference speed by 5x compared to other open reasoning models.
- Leading companies like Microsoft, SAP, and ServiceNow are collaborating with NVIDIA on these models.
- The Llama Nemotron models come in Nano, Super, and Ultra sizes, optimized for different deployment needs.
- NVIDIA conducted extensive post-training using synthetic data and curated datasets to refine the models.
- Microsoft is integrating Llama Nemotron models into Azure AI Foundry to enhance services like Azure AI Agent Service.
- SAP is using the models to improve SAP Business AI solutions and its AI copilot, Joule.
- ServiceNow is leveraging the models to build more accurate AI agents for enterprise productivity.
- Accenture and Deloitte are incorporating the models into their platforms for industry-specific AI solutions.
- NVIDIA AI Enterprise provides tools like the AI-Q Blueprint and AI Data Platform for deploying agentic AI.
- The Llama Nemotron Nano and Super models are available for free via the NVIDIA Developer Program.
I've Had It with Microsoft
9 months ago
- Microsoft is raising prices on existing customers to fund its AI spending.
- Generative AI is criticized as a con with big promises but little utility.
- Microsoft's deceptive practice involves moving customers to a more expensive plan with AI features without clear communication.
- The author's Office subscription price increased by 46%, with no upfront option to opt out of AI features.
- Only when attempting to cancel did the author discover a cheaper, non-AI plan existed.
- The author expresses disgust at Microsoft's deceptive tactics and plans to leave its ecosystem.
- Criticism extends to Microsoft's broader unethical practices, including aiding Israel's actions against Palestinians.
- The author calls for consequences for Microsoft and similar tech companies deceiving customers.
Trump's AI plan says a lot about open source – but here's what it leaves out
9 months ago
- Trump's AI Action Plan emphasizes open-source AI models to maintain US technological dominance.
- The plan encourages freely available open-source AI models to foster innovation globally.
- Open-source AI is seen as a counter to China's growing influence in AI.
- Funding for the National AI Research Resource (NAIRR) is limited, with only $30 million from Biden and tech donations.
- The plan lacks concrete financial support and has many restrictions for federal funding.
- Open-source AI is framed as a strategic tool to shape international AI norms.
- The plan does not mandate open-source releases, leaving decisions to developers.
- Marco Rubio oversees the initiative despite lacking technical expertise in AI or open-source.
Claude Code Introduces Specialized Sub-Agents
9 months ago
- Sub agents in Claude Code are specialized AI assistants for task-specific workflows.
- They have their own context window, tools, and custom system prompts.
- Benefits include context preservation, specialized expertise, reusability, and flexible permissions.
- Sub agents can be created at project or user level, with project-level taking precedence.
- Configuration includes name, description, and optional tool access.
- Management can be done via the `/agents` command or direct file editing.
- Best practices include starting with Claude-generated agents, focusing on single responsibilities, and limiting tool access.
- Advanced usage includes chaining sub agents and dynamic selection based on context.
Anthropic Faces Potentially "Business-Ending" Copyright Lawsuit
9 months ago
- Anthropic faces a potentially 'business-ending' copyright lawsuit over the use of pirated books in training its AI models.
- A federal judge certified a class action lawsuit, the first of its kind in the generative AI context, which could result in damages exceeding a billion dollars.
- Anthropic's use of pirated books was ruled as copyright infringement, with the jury to decide the extent of damages.
- The company is fundraising at a valuation potentially over $100 billion, but much of its funding is in cloud credits, not cash.
- Legal experts warn that the lawsuit could lead to 'business-ending liability' for Anthropic.
- Anthropic argued that the court failed to account for the challenges of establishing ownership for millions of books in a single lawsuit.
- The case highlights the fine line between 'fair use' of copyrighted materials for AI training and outright copyright infringement.
- Anthropic's handling of pirated data, including keeping it accessible for internal use, has not helped its case.
- The company is also facing a separate copyright lawsuit from major music publishers over reproduced lyrics.
- Anthropic's legal troubles could set a precedent affecting other AI companies like OpenAI and Meta.
- The industry is pushing for regulatory clarity on the use of copyrighted materials in AI training.
- Anthropic's CEO has signaled a willingness to accept investment from autocratic regimes to secure necessary funding.
- The outcome of the lawsuit could have significant implications for the AI industry's approach to copyright and data use.
Upsides and Downsides
9 months ago
- Startups must shift their product approach as they grow, moving from focusing on 'upsides' (strong-link problems) to 'downsides' (weak-link problems).
- Early-stage startups thrive on solving strong-link problems where variance and outliers are beneficial.
- Mature startups face weak-link problems, requiring reliability, security, and risk minimization to satisfy late adopters.
- Founders from big companies often struggle with early-stage startups because they don't prioritize exploration over perfection.
- Balancing core product maintenance with new, high-variance bets is crucial for mature companies.
- Segment used the McKinsey horizon framework: 60% core, 30% emerging bets, 10% new bets.
- AI products currently excel in the 'upside' phase but struggle with reliability needed for mass adoption.
- Successful AI tools like Copilot and Midjourney cross the chasm by being reliable, not just impressive.
- Regularly assess whether customers prioritize upside or downside to adjust product focus accordingly.
UN chief urges AI companies to focus on renewable energy
9 months ago
- UN Secretary-General Antonio Guterres urges AI sector to prioritize sustainable energy for data centers.
- A typical AI data center consumes as much electricity as 100,000 homes.
- AI data center energy demand is rapidly expanding, potentially matching Japan's total electricity consumption by 2030.
- Guterres calls for major tech firms to power all data centers with 100% renewables by 2030.
- AI can help improve energy efficiency but is also highly energy-intensive.
- Data centers currently account for 1.5% of global electricity consumption, expected to more than double by 2030.
Doge is suggesting an AI tool that puts half of federal regs on a 'delete list'
9 months ago
- DOGE is developing an AI tool called the 'DOGE AI Deregulation Decision Tool' to analyze federal regulations.
- The tool identified 100,000 out of 200,000+ regulations as candidates for deletion, aiming to cut 50% of federal regulations.
- Agencies are expected to complete deregulation lists by September 1, with the tool promising a four-week turnaround.
- The AI targets regulations no longer required by law, with human staff reviewing deletions before finalization.
- The tool has been tested at the CFPB (used for 100% of deregulations) and HUD (1,083 regulatory sections reviewed).
- HUD employees reported occasional inaccuracies, including misinterpreting legal language.
- DOGE plans to train other agencies on the tool starting this month.
New AI architecture is 100x faster in reasoning with just 1k training examples
9 months ago
- Sapient Intelligence, a Singapore-based AI startup, developed the Hierarchical Reasoning Model (HRM), a new AI architecture that outperforms large language models (LLMs) in complex reasoning tasks while being smaller and more data-efficient.
- HRM is inspired by the human brain's use of distinct systems for slow, deliberate planning and fast, intuitive computation, achieving impressive results with less data and memory than LLMs.
- Current LLMs rely on chain-of-thought (CoT) prompting, which has limitations like brittleness and high data requirements, leading to long, slow responses.
- HRM uses 'latent reasoning,' reasoning internally without explicit language generation, similar to human thought processes, avoiding issues like vanishing gradients and early convergence.
- HRM consists of two coupled modules: a high-level (H) module for abstract planning and a low-level (L) module for fast computations, enabling deep reasoning without extensive data.
- HRM outperformed state-of-the-art CoT models on benchmarks like ARC-AGI, Sudoku-Extreme, and Maze-Hard, achieving near-perfect accuracy with minimal training data.
- HRM's efficiency offers significant cost savings and speedups, making it suitable for latency-sensitive fields like robotics and data-scarce domains like scientific exploration.
- Sapient Intelligence is evolving HRM into a general-purpose reasoning module, with promising applications in healthcare, climate forecasting, and robotics.
Everything I've Learned so far About OpenAI's Agents
9 months ago
- OpenAI's new 'Agents' product is receiving significant hype despite being unfinished and not fully functional.
- The author tested Agents extensively, finding it largely ineffective for tasks like creating PowerPoints and online shopping.
- Agents include tools like a Chromium-based browser, virtual desktop, Linux environment, and image generation, but many features are unreliable.
- Core functions include GUI interactions, Python programming, file operations, and web development, but with frequent errors and limitations.
- Agents cannot install new applications, struggle with authenticated websites, and produce poor-quality outputs like poorly formatted documents.
- Despite current flaws, the author believes AI agents are the future and will be pushed by major tech companies like Google and Microsoft.

Hasty Briefsbeta

#ai

You Can Now Disable All AI Features in Zed

The Last 100k Hacker News Links

Human teens beat AI at an international math competition

How big tech is force-feeding us AI

White House unveils plan to "win" global AI race through deregulation

You can now train a 70B language model at home

Anthropic teams use Claude Code

Qwen3-235B-A22B-Thinking-2507

Google Opal

Efrit: A native elisp coding agent running in Emacs

Nvidia Launches Family of Open Reasoning AI Models: OpenReasoning Nemotron

I've Had It with Microsoft

Trump's AI plan says a lot about open source – but here's what it leaves out

Claude Code Introduces Specialized Sub-Agents

Anthropic Faces Potentially "Business-Ending" Copyright Lawsuit

Upsides and Downsides

UN chief urges AI companies to focus on renewable energy

Doge is suggesting an AI tool that puts half of federal regs on a 'delete list'

New AI architecture is 100x faster in reasoning with just 1k training examples

Everything I've Learned so far About OpenAI's Agents