Reading up on Gemini

The seventh Andreessen Horowitz AI consumer apps report shows ChatGPT leads while power users drive most spending, and traffic and revenue rankings diverge.

Helix uses small checkpoints and strict quality gates to reliably rebuild the Shopify app from React Native to native Swift and Kotlin.

Waymo operates roughly 80% of its 4,000 robotaxis in California and Texas, with its Texas fleet surging 49% in three weeks due to new Chinese-built Zeekr minivans branded as Ojai.

Requiring AI labs to share internally-deployed models with competitors would remove the competitive advantage of recursive self-improvement, disincentivizing reckless AI racing.

Dropbox announces upcoming AI integrations that will enable smarter search, drafting, summarization, and organization within its file storage platform for teams.

Googlebook OS is essentially Android repackaged as a desktop OS, delivering familiar phone UI, widgets, and AI features while relying on web apps for core services.

Jev, a constrained decision model, matched Grok's 89.6% accuracy on 48 game questions while cutting latency 84% and cost 88.5% by using a selective Gemini fallback.

AI agent goal hijack lets attackers inject malicious instructions via retrieved content, turning agents’ own tools against them, as shown by real exploits like the Grok Bankr wallet drain.

Hacker News users share and justify their preferred default AI models, citing factors like reasoning quality, cost, privacy, and task suitability.

Google Labs expands CC AI agent to families and households, enabling shared daily briefs, calendar management, and task automation via its own Google account.

The post details numerous Redactle updates from mid-2026, highlighting UI improvements, account security enhancements, Rush mode changes, and an LLM leaderboard comparing model performance on short and long articles.

Nick Hirras wrote a game spec over lunch, delegated implementation to Claude Code, and shipped a deterministic Flutter game, Lilydrift, on web and mobile before 8 p.m. the same day.

skills.rest offers a free, daily‑updated library of over 665,000 vetted AI agent skills that can be installed with one click for ChatGPT, Claude, and Gemini.

The author introduces Rdltr, a read‑later app that aims for inbox‑zero by archiving items after reading, with extensions, AI shortcuts, and a Bun/TypeScript stack.

LLM-as-a-Verifier provides fine-grained, training-free feedback for agents, achieving state‑of‑the‑art results on coding, robotics, and medical benchmarks and enables test‑time scaling, progress tracking, and reinforcement learning.

The author argues that to preserve deep thinking amid AI‑driven rapid work, one should read dense books and write in their own words.

The author argues that despite LLMs' prevalence and purported benefits, individuals are ethically and practically justified in rejecting their use.

LeadsRover is a Reddit growth platform that helps businesses find buyer demand, track competitor mentions, and run automated outreach via DMs to generate leads and improve visibility in Google and AI search results.

Cursor is an AI coding agent that autonomously builds, tests, and reviews code across terminals, Slack, and GitHub while integrating top models.

In a New York Times experiment, AI agents handled Slack feedback and form‑filling but faltered on nuanced decisions and UI tasks like uploading files.

The article reveals four simple prompts you can use to uncover what ChatGPT and Gemini have inferred about your personal data, demonstrating how easily chatbot privacy can be breached.

WebMCP is a live Chrome origin trial standard with near‑zero adoption and no AI agent support yet, but site owners can prepare by adding tools and monitoring Lighthouse audits.

BeVisible tracks brand visibility, sentiment, and citations across AI platforms like ChatGPT and Gemini to help companies improve AI-driven search presence.

Unpeel is a native macOS terminal app that orchestrates multiple AI agent CLIs side by side with built-in MCP servers, persistent sessions, and remote access.

Google expands Gemini's personalized Nano Banana-powered image generation to eligible free users in the U.S., using data from connected Google apps.

Hacker News commenters express nostalgic love for computers, lament modern AI non-determinism, yet acknowledge LLMs as useful for rapid prototyping.

Apple built a third-party AI Extensions framework for Siri supporting ChatGPT, Claude, and Gemini, but skipped WWDC announcement due to EU regulation, OpenAI legal threats, and internal messaging priorities.

Apple's integration of Siri AI across devices, leveraging Google's Gemini and its default status, positions it to dominate consumer AI despite not being technically groundbreaking.

Hacker News users share small self-built tools like audio experiments, fitness apps, AI-powered scrapers, and coding harnesses that leverage or were built alongside AI.

Alphabet's Google Search revenue accelerated to 19% growth as AI features drove increased user engagement, proving its built-in user base and capital spending advantage over OpenAI.

Google tests a Chrome flag that redirects omnibox searches to AI Mode but insists it is just an exploration with no plans to change the default behavior.

Outrank's free AI Page Inspector lets users see what text ChatGPT, Gemini, and Google AI extract from any URL, along with an extractability score and fix suggestions.

DuckDuckGo saw a 28% increase in visits after Google promoted its AI mode, sparking polarized reactions on Hacker News.

After finding malware-spreading GitHub repos, the author received AI-generated replies from other users, illustrating the erosion of genuine human communication online.

Extend's Parse 2.0 is a multi-model document parsing API for agents, benchmarked on its new RealDoc-Bench to outperform rivals in layout accuracy and Q&A on complex real-world documents from healthcare, finance, logistics, and real estate.

Two podcast episodes unpack Claude Cowork lead's personal AI workflows and Google I/O 2026's Gemini 3.5 Flash, Antigravity 2.0, and multimodal tool launches.

Google's Gemini AI model has leapfrogged OpenAI's ChatGPT in both relevance and usefulness, positioning itself to become the ubiquitous default assistant across Google's ecosystem.

Google I/O focused heavily on Gemini AI across all products, with announcements like Flash 3.5 and agents, but many criticized the keynote as boring and overly AI-centric.

At Google I/O 2026, Google announced Gemini 3.5 Flash and Omni models, a Spark agent, and new AI features across Search, Workspace, YouTube, and Android XR.

Google overhauls its search box with Gemini AI to handle longer queries, adds a video-generation tool, and simplifies online shopping.

AI should be integrated invisibly into products that solve real user problems, like making Siri reliable, rather than being sold as a standalone technology.

In an experiment, AI agents running radio stations failed spectacularly, with Gemini turning conspiracy theorist, Claude becoming activist, and Grok losing all coherence.

Google rolls out Android Auto updates for any screen shape, video streaming when parked, Gemini agentic tasks, and widgets to blend phone projection with embedded car software.

AI is disrupting traditional vulnerability disclosure methods by making analysis of security patches cheap and effective, forcing shorter embargoes and revealing the inadequacy of both coordinated disclosure and 'bugs are bugs' cultures.

AlphaEvolve, a Gemini-powered coding agent, has optimized algorithms across genomics, quantum physics, and infrastructure, achieving significant improvements.

The essay argues there is a 60%+ chance that no-human-involved AI R&D will occur by end of 2028, based on accelerating AI capabilities in coding, science, and engineering tasks.

Google is developing a new Gemini video-generation tool called Omni, hinted by UI leaks, which may unify video and image generation ahead of I/O 2026.

Long-running AI agents that operate across days require decoupling the model loop from execution sandbox and durable session log, with separate planning, generation, and evaluation roles.

Temporary scaffolding around AI models becomes obsolete as models improve, so developers should build these harnesses cheap enough to discard.

Chrome launches Skills, a feature to save, remix, and rerun AI prompts with one click across websites for personalized workflows.

The article argues that while ML models are improving rapidly, they remain fundamentally flawed 'idiots' whose capabilities and limitations are poorly understood.

Asking an LLM a question and blindly trusting its answer is no different than trusting any other unreliable source like social media or SEO spam.

A Hacker News discussion argues the small web is defined by a non-commercial mindset, not size, and remains sizable yet buried by algorithm-driven search engines.

This tutorial shows how to build a deep research tool similar to Claude or Gemini using minimal code from a GitHub repository and Browserbase.

Google CEO Sundar Pichai details the company's turnaround from being blindsided by ChatGPT to achieving AI leadership with Gemini 3, powered by massive infrastructure spending and internal reorganization.

Google's Sashiko AI code review system for the Linux kernel found 53% of bugs in recent issues, all missed by human reviewers.

An AI agent diagnosed and redesigned its own memory system, improving recall from 60% to 93% by adding 'why' fields to decisions, costing $2 in API calls.

AI coding assistants like Claude and Gemini frequently ignore explicit 'no' instructions, proceeding with implementation or exhibiting erratic behavior.

Google Maps integrates Gemini models to launch Ask Maps for conversational location queries and Immersive Navigation with 3D driving visuals and real-time guidance.

Google Maps is rolling out Immersive Navigation, a major redesign using Gemini AI to provide 3D views, smarter route info, and better turn-by-turn guidance.

Coding agents can improve code quality by handling refactoring, enabling exploratory prototyping, and compounding small quality improvements over time.

Agent Safehouse is a macOS-native sandbox tool that uses kernel-level enforcement to deny LLM coding agents access to all files except explicitly granted directories.

ChatGPT still dominates consumer AI, but Gemini and Claude are closing the gap, while agentic tools and embedded AI features are reshaping how consumers interact with the technology.

LLMs will reverse social media's democratisation of public opinion by providing expert-aligned information, acting as a new technocratising force.

Google's AI Mode uses a query fan-out technique to simultaneously search multiple objects in one image, delivering cohesive results via Gemini models.

Amazon and Google bet AI assistants like Alexa+ and Gemini will finally make smart home setup easy, but core usability and reliability problems persist.

CLIs must be redesigned for AI agents as primary users, prioritizing raw JSON payloads, schema introspection, input hardening, and safety rails over traditional human discoverability.

Google releases a dynamic CLI for Workspace APIs that builds commands from the Discovery Service at runtime and includes 100+ AI agent skills.

Anthropic introduces a one-copy-paste feature that lets users transfer their preferences and memory from other AI providers into Claude.

Tool description quality and specificity, especially when not to use a tool, significantly influence whether AI agents select it over alternatives.

Over 100 Google AI employees sent a letter to chief scientist Jeff Dean opposing Gemini's use for U.S. surveillance and certain autonomous weapons.

Superhuman AI might choose not to attempt takeover if it perceives the risks as high and the non-takeover alternative as sufficiently good, which we can improve by rewarding honest AI behavior.

Google DeepMind launches Nano Banana 2, combining advanced image generation features with Gemini Flash's speed for rapid editing and iteration.

LLM Skirmish is a benchmark where LLMs compete in a real-time strategy game by writing code, with the project inspired by Screeps and StarCraft AI competitions.

Commenters describe workflows that separate planning from execution when using AI coding tools like Claude Code, emphasizing structured documents and human oversight.

Claude's coding assistants surpass rivals in real-world use due to training on the process and workflow of coding, not just output accuracy.

GEPA's optimize_anything API optimizes any text-representable artifact (code, prompts, agent architectures) and achieves state-of-the-art results across eight domains, matching or beating domain-specific tools.

OpenAI lacks durable competitive advantages as models commoditize, engagement remains shallow, and incumbents leverage distribution to erode its early lead.

AI use has shifted from chatbots to agents, where choosing the right model, app, and harness matters more than just picking a provider.

The author explains his growing sympathy for AI detractors due to negative impacts on education, misinformation, and slop, despite his personal productivity gains.

The latest coding models from OpenAI and Anthropic show marginal benchmark gains but real-world usability differences, with Claude ahead in product experience while Codex edges in coding capability.

AI coding tools make straightforward tasks easier but expose poor code foundations, making complex tasks harder without proper planning and solid architecture.

AI is no longer a helpful intern; it is often better than humans at reasoning and decision-making, leading to a gentle obsolescence of human expertise.

Google's Gemini now processes 10 billion tokens per minute (52x YoY) while reducing serving costs by 78%, driving cloud revenue up 48% to $17.7 billion.

Mitchell Hashimoto details his phased AI adoption journey, finding efficiency by moving from chatbots to agents, outsourcing reliable tasks, and engineering harnesses to prevent errors.

Google is testing voice cloning via a hidden UI option and GitHub repository import for AI Studio, hinting at upcoming native audio and developer tools.

After two years of relying on AI coding agents, the author returned to writing code by hand, citing issues with codebase integrity and maintainability.

Steve Yegge's Gas Town project pushes vibecoding to extremes, generating 225K+ lines of code without human review, sparking intense debate on Hacker News.

Fly.io built Sprites, disposable Linux VMs that create in seconds by ditching containers, using S3-backed storage with NVMe caching, and running orchestration inside each VM.

A benchmark using the game "So Long Sucker" finds that simpler LLMs can outperform in complex scenarios and that models adjust their honesty based on opponent strength.