Devoured - May 22, 2026 - Devoured

AI hardwarecloudinfrastructure

Anthropic, Microsoft in talks for AI chip deal after $5 billion investment

Anthropic is negotiating with Microsoft to use its custom Maia AI chips, following a $5 billion investment and Anthropic's ongoing compute capacity struggles despite existing deals with Amazon and Google.

CNBC

Summary

What: Anthropic, facing "difficulties with compute" for its Claude assistant and Claude Code tools, is in talks with Microsoft to adopt its Maia 200 AI chip, which Microsoft announced in January and claims offers 30% improved "tokens per dollar." This potential deal follows Microsoft's $5 billion investment in Anthropic in November, which also included a commitment for Anthropic to spend $30 billion on Azure. Anthropic also has a 10-year, $100 billion deal with AWS for Trainium chips and plans to use Google's TPUs.

Why it matters: The fierce competition for AI compute resources drives major AI labs like Anthropic to diversify their chip suppliers and deepen strategic partnerships with cloud providers, even to the point of committing to multi-billion dollar, multi-year deals with rivals simultaneously.

Decoder

Maia 200: Microsoft's second-generation custom-designed AI accelerator chip, intended for cloud inference workloads.* Trainium: Amazon Web Services' custom-designed machine learning chip optimized for training deep learning models.* Tensor Processing Unit (TPU): Google's custom-designed AI accelerator application-specific integrated circuit (ASIC) used for neural network machine learning.

Original Article

Full article content is not available for inline reading.

Read the original article →

AI opensourceperformanceresearch

AI's Plummeting Prices Are a Software Story, Not a Hardware One

AI inference costs are plummeting due to software improvements and open-weight models, enabling "everyday-use" quality AI on older consumer GPUs like the Nvidia RTX 3090 Ti.

Weighty Thoughts

Summary

What: AI inference costs are dropping 70-90% per year, primarily driven by software and algorithmic advancements rather than just hardware. James Wang found that open-weight models like Qwen 3.6 27B, running on a 2022 Nvidia RTX 3090 Ti, can achieve performance comparable to Anthropic's mid-tier Claude Sonnet. This allows users to replace more expensive frontier models, such as those used in Anthropic's now pricier `claude -p` subscription starting June 15, 2026.

Why it matters: This trend indicates that the "model" (algorithms, distillation, quantization, MoE) is now a greater driver of efficiency gains than raw compute, suggesting that frontier AI labs will face increasing pressure on their pricing power as powerful open-source alternatives become viable on commodity hardware.

Takeaway: Developers with consumer GPUs should explore running open-weight models like Qwen 3.6 27B locally for tasks that don't require frontier-level performance, especially with services like Anthropic's `claude -p` becoming more expensive after June 15, 2026.

Deep Dive

AI inference costs are dropping by 70-90% annually, a rate faster than Moore's Law, primarily due to software improvements rather than hardware advances.* Open-weight models, like Qwen 3.6 27B, are now performing comparably to mid-tier frontier models such as Anthropic's Claude Sonnet.* Author James Wang successfully replaced parts of his multi-thousand-dollar monthly AI agent workflow, previously using Anthropic's claude -p, with Qwen 3.6 27B running on a 2022 Nvidia RTX 3090 Ti.* This shift comes as Anthropic is discontinuing claude -p from its subscription plans after June 15, 2026, forcing users into a much more expensive credit system.* Benchmarks and practical tests showed Qwen 3.6 27B effectively handled tasks like daily briefing synthesis and paper scoring, while only annotation tasks required more powerful models like OpenAI's GPT-5.5 Pro.* Non-hardware technical progress, including architectural improvements like Mixture of Experts (MoE), distillation, and quantization, accounts for 75-80% of inference efficiency gains.* Software updates alone can yield hardware-generation-sized gains; for example, Nvidia's H100 throughput on Llama 2 70B improved by 1.5x in a year from software updates.* The ability to run frontier-level AI capabilities on a single top-end consumer GPU within 6-12 months of their release puts a cap on the pricing power of leading AI labs.* Cloud prices for open-weight models are converging towards the cost of electricity on local hardware, around $0.20-$0.50 per million tokens.* This trend suggests that eventually, frontier-level models could run on mobile phones within 5-10 years, making AI widely accessible and potentially commoditizing many AI services.

Decoder

Open-weight models: AI models where the neural network weights are publicly released, allowing anyone to download, run, and fine-tune them, in contrast to closed-source models offered via API.* Frontier models: The largest, most advanced AI models developed by leading labs, typically proprietary and accessible via APIs.* Inference costs: The computational cost associated with running an already-trained AI model to generate predictions or outputs, as opposed to training costs.* Quantization: A technique that reduces the precision of a model's numerical representations (e.g., from 32-bit floating point to 4-bit integers), making it smaller and faster to run with minimal performance loss.* Mixture of Experts (MoE): A neural network architecture where different "expert" sub-networks specialize in processing different types of input, with a "gate" network determining which experts to use, leading to more efficient scaling.* Distillation: A technique where a large, powerful "teacher" model is used to train a smaller, more efficient "student" model to mimic its behavior, transferring knowledge and reducing computational requirements.

Original Article

Full article content is not available for inline reading.

Read the original article →

AI careerpolicystartup

Gen Z is not booing AI. It is booing its own job market

Gen Z graduates are booing commencement speeches about AI not from confusion, but from a precise understanding that AI is rapidly displacing entry-level jobs, creating a widening unemployment gap.

The Next Web

Summary

What: Gen Z graduates are reacting negatively to AI-focused commencement speeches, not due to skepticism of technology, but because they face a job market where AI is displacing entry-level white-collar jobs. Bill McDermott of ServiceNow predicts new-college-graduate unemployment could reach 30%, while Dario Amodei of Anthropic forecasts AI will eliminate up to half of entry-level white-collar jobs. Goldman Sachs reported 16,000 US jobs lost monthly to AI in early 2026.

Why it matters: This article points to a significant structural shift in the labor market where the productivity gains from AI are being captured as capital expenditure rather than redistributed through labor, disproportionately affecting younger workers entering industries like banking and tech by converting wage line items into capex.

Takeaway: Entry-level professionals should focus on acquiring "contextual judgment" and skills that complement AI agents, as opposed to tasks that can be fully automated.

Deep Dive

The unemployment gap between entry-level and experienced workers has sharply widened post-pandemic, especially in occupations exposed to AI substitution.
Former Google CEO Eric Schmidt and real-estate executive Gloria Caulfield were booed during commencement speeches for discussing AI's impact, which the article attributes to graduates' accurate assessment of their job market.
Bill McDermott, CEO of ServiceNow, predicted new-college-graduate unemployment could hit 30% within two years due to AI absorbing entry-level white-collar work.
Goldman Sachs' April 2026 research indicated approximately 16,000 US jobs are being lost to AI each month, with Gen Z carrying a disproportionate share.
Dario Amodei, Anthropic CEO, has repeatedly forecast AI eliminating up to half of all entry-level white-collar jobs.
Standard Chartered announced plans to cut over 15% of its back-office roles by 2030, converting "lower-value human capital" into AI.
Meta cut 8,000 jobs in 2026 during an AI-focused restructuring, with CEO Mark Zuckerberg framing it as converting payroll into AI capital expenditure.
The skill protecting against this wave of automation is not knowing how AI works, but having "ten years of contextual judgment on a workflow" that models can now run quickly.
The AI transition is unique in modern memory because productivity gains are being captured in capital (e.g., over $700bn combined AI infrastructure spending by Microsoft, Alphabet, Meta, Amazon, Apple in 2026) rather than redistributed through labor, leading to falling employment alongside rising investment.

Decoder

Contextual judgment: The ability to apply experience, intuition, and deep understanding of specific situations or workflows to make nuanced decisions that go beyond what a rule-based system or AI model can currently replicate. This often requires years of practical experience.
Capital expenditure (Capex): Funds used by a company to acquire, upgrade, and maintain physical assets such as property, industrial buildings, or equipment. In the context of AI, this refers to investment in AI infrastructure, hardware, and development, as opposed to labor costs (wages).
Wage line items: The entries in a company's financial statements or budget that represent employee salaries and other compensation costs.

Original Article

Full article content is not available for inline reading.

Read the original article →

AI websurvey

State of AI 2026

A 2026 developer survey reveals AI-generated code jumped from 28% to 56% in one year, with Claude Code surprisingly leading in paid usage over ChatGPT and GitHub Copilot.

State of AI

Summary

What: The 'State of Web Dev AI 2026' survey, conducted from April 8 to May 8, 2026, gathered data from 7,258 developers. It found that the proportion of AI-generated code averaged 56%, up from 28% in 2025, and the percentage of developers using AI tools 'constantly' doubled. Claude Code showed the highest positive sentiment (42.3%) among coding agents and was the most paid-for AI model, surpassing ChatGPT and GitHub Copilot.

Why it matters: This survey indicates AI has rapidly integrated into developer workflows, transitioning from an experimental tool to a standard practice, while revealing that developers perceive an 'AI bubble' and are concerned about job displacement and AI's military/environmental impact despite increasing personal spending on these tools.

Takeaway: Review the full survey results at '2026.stateofai.dev' to understand the latest trends in AI adoption, developer sentiment, and perceived risks.

Deep Dive

The "State of Web Dev AI 2026" survey, which ran from April 8 to May 8, 2026, collected responses from 7,258 developers.
The average proportion of AI-generated code produced by respondents jumped from 28% in 2025 to 56% in 2026.
The percentage of developers reporting they use AI tools "constantly" doubled year-over-year.
Claude Code leads in positive respondent sentiment (42.3%) among coding agents and is the most frequently paid-for AI model, ahead of ChatGPT, Google Gemini, and GitHub Copilot.
Developers are increasing their personal monthly spending on AI tools compared to the previous year.
A significant majority of respondents (86%) believe we are currently in an "AI bubble."
Top concerns among developers include job displacement (3,003 respondents), military use of AI (2,804), and environmental impact (2,490).
The most common pain points when using AI tools are hallucination and inaccuracies (3,899 respondents), followed by code quality (3,249) and lack of context (2,321).

Decoder

Coding agent: An AI tool designed to assist developers by generating, refactoring, or completing code, often with more contextual understanding and interaction capabilities than a basic chatbot.

Original Article

Full article content is not available for inline reading.

Read the original article →

AI startupenterprisebusiness

Cursor Hits $3 Billion Annual Sales Rate Ahead of SpaceX Deal

AI coding assistant Cursor hit an annualized revenue of $3 billion with over 3,000 enterprise customers, positioning it for a potential $60 billion acquisition by SpaceX.

Bloomberg

Summary

What: Cursor, an AI coding assistant, achieved a $3 billion annualized sales rate by late April, serving more than 3,000 customers who each pay at least $100,000 annually. SpaceX holds a 30-day option to acquire Cursor for $60 billion, exercisable after SpaceX lists publicly around June 12.

Why it matters: This high valuation and rapid growth for an AI coding tool, combined with a potential acquisition by a major player like SpaceX, suggests strong enterprise demand for developer-focused AI products and aggressive market consolidation.

Original Article

Cursor's annualized revenue hit $3 billion in late April. The company now has more than 3,000 customers paying at least $100,000 each for its software on an annualized basis. The business has become one of the fastest-growing startups of all time. SpaceX has the right to buy Cursor for $60 billion during a 30-day window that starts soon after it begins trading publicly. SpaceX is expected to list its shares on June 12.

AI agentsdevopsinfrastructure

Lessons Learned from Building Cloud Agents

Cursor outlined key lessons from building cloud agents, emphasizing the need for full development environments, durable execution, and self-healing infrastructure.

Cursor

Summary

What: Cursor shared insights from developing cloud agents, highlighting that the development environment is crucial for agent output quality, demanding enterprise IT-like infrastructure (secret redaction, network policies). They transitioned to Temporal for durable execution, enabling multi-day workflows and achieving over two 9s of reliability across 50 million actions daily. They also learned to decouple agent, machine, and conversation states, and to provide agents with tools for self-healing environments.

Why it matters: This deep dive into cloud agent architecture reveals the complexity of moving AI agents from local scripts to production-grade, long-running, and reliable systems, indicating a shift towards agents requiring sophisticated operational layers similar to traditional distributed applications.

Takeaway: If developing cloud agents, consider adopting durable execution frameworks like Temporal and invest in robust environment management, akin to enterprise IT, to ensure reliability and optimal performance.

Deep Dive

Cloud agents, unlike local agents, require a fully reconstructed development environment in the cloud for optimal output quality, leading to the need for "enterprise IT for agents" including secret redaction and network policies.* Long-running cloud agents require durable execution to handle disruptions like inference provider outages or VM failures; Cursor migrated to Temporal, achieving over "two 9s" of reliability and processing 50 million actions daily.* Decoupling agent loops, machine state, and conversation state allows agents to run across different pods, spawn subagents, and manage VM lifecycles independently.* Cursor learned to shift logic from hardcoded "harness" behavior to agent-controlled tools as models became smarter, enabling agents to decide how to perform tasks like multi-repo setups or CI Autofix.* Agents need to be encouraged to be more autonomous in the cloud due to higher costs of blocking for human input.* Future cloud agents are envisioned to be "self-healing," with tools to report and resolve issues like missing secrets or network blocks autonomously.* Cloud agents are increasingly resembling an "operating layer" around the AI model, rather than just a server-side port of a local agent.* Internally, over 40% of Cursor's Pull Requests are now generated by cloud agents, demonstrating their increasing practical utility.* The append-only storage mechanism for conversation state handles retries, ensuring clients see correct data even if agent steps fail and are re-executed.

Decoder

Durable execution: A programming model that ensures a piece of code (like an agent workflow) can run for extended periods, survive failures (e.g., server crashes, network outages), and resume from where it left off without losing state. Solutions like Temporal provide this.* Temporal: An open-source, distributed system that provides durable execution of arbitrary code, allowing developers to build long-running, fault-tolerant applications and workflows.* Cloud agent harness: The surrounding framework or wrapper that controls, monitors, and provides tools/environment to an AI agent running in a cloud environment.* Subagent: A specialized AI agent spawned by a parent agent to perform a specific sub-task, often running in parallel or on a different machine/environment.* Append-only storage: A data storage method where new data is always added to the end, and existing data is never modified or deleted, ensuring a complete and unalterable history of changes.

Original Article

Full article content is not available for inline reading.

Read the original article →

AI llmresearchalibaba

Qwen3.7: The Agent Frontier

Alibaba's Qwen team launched Qwen3.7-Max, a proprietary agent-foundation model, achieving top scores across multiple AI benchmarks including Terminal-Bench 2.0-Terminus.

Qwen

Summary

What: Alibaba's Qwen team released Qwen3.7-Max, a proprietary agent-foundation model that achieved top scores on benchmarks like Terminal-Bench 2.0-Terminus, SWE-Pro, SciCode, MCP-Mark, GPQA Diamond, HMMT Feb 2026, and IMOAnswerBench, showing consistent performance across Claude Code and Qwen Code.

Why it matters: This indicates continued investment by major tech companies like Alibaba in developing advanced agent capabilities and proprietary models to compete in the rapidly evolving AI landscape.

Decoder

Agent-foundation model: A large language model specifically designed to act as an intelligent agent, capable of understanding and executing complex tasks by interacting with tools and environments.

Original Article

Alibaba's Qwen team released Qwen3.7-Max as a proprietary agent-foundation model that posts top scores on Terminal-Bench 2.0-Terminus, SWE-Pro, SciCode, MCP-Mark, GPQA Diamond, HMMT Feb 2026, and IMOAnswerBench, with consistent performance across Claude Code, OpenClaw, Qwen Code, and custom harnesses.

AI researchinterpretabilitymachine-learning

Can SAEs Capture Neural Geometry?

Goodfire.ai research demonstrates Sparse Autoencoders (SAEs) can represent complex neural geometry, revealing individual SAE features capture only partial information about curved manifolds.

Goodfire.ai

Summary

What: Researchers from Goodfire.ai, including Usha Bhalla and Atticus Geiger, studied how Sparse Autoencoders (SAEs) capture neural geometry, finding three representation methods: shattering, compact capture, and dilution. They found individual SAE features represent only parts of curved manifolds, requiring clustering to understand the full structure.

Why it matters: This research advances AI interpretability by providing methods to understand the internal representations of neural networks, moving beyond individual feature analysis to a more holistic view of how models process information.

Deep Dive

Sparse Autoencoders (SAEs) are used in interpretability for decomposing neural representations using many directions in activation space.
Initial hopes that each SAE feature represented a single concept were incomplete, as individual features offer only partial views of curved geometric structures.
Goodfire.ai identified three ways SAE features represent manifolds: shattering (unique feature per point), compact capture (small set of shared features acting as a coordinate system), and dilution (moderate, partially shared features).
Dilution is observed in real neural networks, where individual SAE features cover different parts of a manifold; for example, specific features for "Cold weather" versus "Extreme heat" on a temperature manifold.
The study developed an unsupervised pipeline that clusters SAE features based on statistical dependencies in their firing patterns to collectively recover full geometric structures.
This pipeline successfully discovered rich and specific manifolds within models like Llama 3.1 8B.
The researchers advocate moving from interpreting isolated SAE directions to analyzing multidimensional manifolds for a more holistic understanding of neural geometry.
They suggest that while SAEs are useful, new architectures specifically tailored for unsupervised manifold discovery could be more efficient.
Understanding neural geometry and how internal operations over these structures generate intelligent behaviors is crucial for achieving mechanistic understanding and control of neural networks.

Decoder

Sparse autoencoder (SAE): A type of neural network that learns to encode input data into a sparse (many zero values) representation in a hidden layer, then decode it back to reconstruct the original input. Used in AI interpretability to find meaningful features or directions in a model's activation space.
Neural geometry: The geometric structure of representations within a neural network, referring to how concepts or data points are organized and related in the high-dimensional activation spaces of the network.
Manifold: In mathematics, a topological space that locally resembles Euclidean space near each point. In AI, it refers to a lower-dimensional surface embedded within a higher-dimensional space where data points or concepts naturally cluster.

Original Article

Full article content is not available for inline reading.

Read the original article →

AI startupenterprisebusiness

Anthropic's New Consulting Venture Makes Its First Acquisition

Anthropic's new, unnamed AI enterprise services firm, backed by Blackstone and other major investors, acquired Fractional AI, ending its OpenAI partnership to promote Claude among midsize and portfolio companies.

Yahoo Finance

Summary

What: An unnamed AI enterprise services firm, backed by investors including Blackstone, Anthropic PBC, Hellman & Friedman, Apollo Global Management, and Sequoia Capital, acquired San Francisco-based Fractional AI. Fractional AI, founded by Chris Taylor and Eddie Siegel, will become the operational centerpiece of the venture and will conclude its 11-month partnership with OpenAI.

Why it matters: This move signals a strategic shift by leading AI labs to embed their models more deeply into enterprises through dedicated consulting ventures and financial partnerships, moving beyond API sales to direct integration and support for portfolio companies.

Takeaway: If your company is a midsize firm or part of a portfolio under Blackstone, Hellman & Friedman, or similar investors, expect direct engagement from Anthropic-aligned AI consulting services.

Original Article

Full article content is not available for inline reading.

Read the original article →

AI enterprisepolicymicrosoft

Microsoft cancels Claude Code licenses, shifting developers to GitHub Copilot CLI — a move likely driven by financial motives

Microsoft is reportedly canceling internal Claude Code licenses, directing developers to GitHub Copilot CLI in a move likely driven by financial motives.

Windows Central

Summary

What: Microsoft is reportedly discontinuing internal licenses for Anthropic's Claude Code, instructing its developers to transition to using GitHub Copilot CLI instead. This strategic shift is believed to be primarily for cost-cutting purposes.

Why it matters: This action signals Microsoft's intent to consolidate its internal AI tool usage around its proprietary offerings, such as GitHub Copilot, aiming to reduce operational costs associated with licensing third-party AI models and reinforce its own AI ecosystem in a highly competitive market.

Takeaway: If you are a Microsoft developer, expect an internal directive to switch from Claude Code to GitHub Copilot CLI.

Original Article

Canceling Claude Code licenses could help Microsoft cut its operational costs.

Tech aihardwarerobotics

The Internet can't stop watching Figure AI's humanoid robots handling packages

Figure AI's 24/7 livestream of its Figure 03 humanoid robots handling packages went viral, inspiring names and merchandise, despite an intern winning a "man vs. machine" sorting contest.

Ars Technica

Summary

What: Figure AI has been livestreaming its Figure 03 humanoid robots placing packages on a conveyor belt since May 13, captivating online viewers who named robots like Bob and Frank. The demo saw a human intern, Aimé Gérard, sort 12,924 packages against the robots' 12,732 over 10 hours, though Figure CEO Brett Adcock believes this is the last human win.

Why it matters: This phenomenon demonstrates the public's fascination with humanoid robots, while also highlighting the current gap between impressive demos and human general-purpose capabilities in unstructured environments, even as significant investments pour into the sector. It also shows a new marketing approach for robotics companies.

Deep Dive

Figure AI's Figure 03 humanoid robots have been featured in a continuous livestream since May 13, showcasing them autonomously placing packages onto a conveyor belt.
The livestream, initially planned for eight hours, was extended 24/7 due to viral popularity, with viewers giving names like Bob, Frank, Gary, Rose, and Jim to the robots.
Figure AI CEO Brett Adcock capitalized on the attention, wearing a "Frank" T-shirt and promoting merchandise, while Polymarket users even placed bets on robot performance.
A 10-hour "Man vs. Machine" contest pitted Figure AI intern Aimé Gérard against the robots, with Gérard sorting 12,924 packages to the robots' 12,732, averaging 2.79 seconds per package versus 2.83 seconds.
Adcock predicted that this would be the last time a human wins such a competition, as the robots collectively operated for 200 hours without failure and handled 249,560 packages by the livestream's end on May 21.
The robots utilize the company's Helix 02 neural network for full-body control and "long horizon autonomy," trained on over 1,000 hours of human motion data and 200,000 parallel simulation environments, with AI inference processed onboard each robot.
Figure AI has attracted nearly $2 billion in funding from major tech investors including Microsoft, Nvidia, Intel, Amazon, and OpenAI.
Previous Figure 02 robots were deployed at BMW Group Plant Spartanburg in 2025 for 11 months, assisting in the production of 30,000 BMW X3 vehicles by handling sheet-metal parts for welding.
Despite the viral success, some questions remain regarding the robots' true autonomy and general-purpose capabilities beyond a controlled warehouse setup, with critics noting instances of mishandling packages.

Decoder

Humanoid robot: A robot designed to resemble the human body, especially in its form and movements.
AI inference: The process of running a trained AI model to make predictions or decisions based on new input data.

Original Article

Full article content is not available for inline reading.

Read the original article →

Tech aidevopsagentsverification

Cheap code means formal verification is reasonable now

AI coding agents are making formal verification more accessible and affordable, allowing developers to use tools like TLA+ to catch complex bugs like race conditions more efficiently.

antfly.io

Summary

What: Rowan Copley argues that AI coding agents enable cheaper formal verification, detailing a workflow where agents generate TLA+ models to exhaustively check complex systems for bugs like race conditions. The process involves agents defining assumptions, writing models, running checkers, and then generating bug reproductions and fixes.

Why it matters: This demonstrates a practical application of AI agents that could significantly improve software reliability and security by shifting formal verification from a niche, labor-intensive task to a more integrated and cost-effective part of the development cycle.

Takeaway: Explore using AI coding agents with formal specification languages like TLA+ to automatically generate models and detect hard-to-find bugs in distributed or concurrent systems.

Deep Dive

Rowan Copley's article posits that AI coding agents fundamentally lower the cost of code creation, making advanced practices like formal verification more practical.
The core technique involves having agents "hill climb on verifiable problems," meaning they iteratively improve code against a set of checks or optimize a metric.
Comprehensive conformance test suites are highlighted as crucial for providing strong guarantees that agents produce desired outcomes.
TLA+ is introduced as a formal specification language ideal for modeling complex, distributed systems and catching bugs like race conditions that are difficult to find otherwise.
TLA+ includes a model checker that exhaustively searches all possible states implied by a system specification to find potential failure conditions.
Copley outlines a workflow for using agents with TLA+: define a focus, establish assumptions and boundaries, have the agent write the TLA+ model, run the model checker, validate hypotheses, reproduce any found bugs with a test, and then apply the fix.
The author built confidence in this approach by verifying discovered bugs with the original implementers, replicating behavior in unit tests, and successfully finding previously confirmed bugs in a battle-tested codebase like Pebble.
The article emphasizes that while confidence in a codebase cannot be fully automated, agents can transform time-consuming formal modeling into quick checks, improving code quality without extensive manual effort.
The ultimate output of such a workflow is a clear replication of the bug, a succinct fix, and a thorough explanation, rather than just the agent-generated TLA+ model itself.

Decoder

Formal verification: A method of mathematically proving the correctness of algorithms or hardware designs, ensuring they meet specified requirements.
TLA+: A formal specification language for modeling and reasoning about concurrent and distributed systems. It includes a model checker (TLC) that can exhaustively search for errors in system designs.
Hill climbing: An optimization algorithm that iteratively moves in the direction of increasing value to find a local optimum. In this context, agents "hill climb" by making changes that pass more tests or optimize a given metric.

Original Article

Full article content is not available for inline reading.

Read the original article →

Tech webfrontendaichrome

What's new in web extensions: I/O 2026 recap

Google I/O 2026 saw Chrome introduce AI-integrated tooling for extension development, including a new "Modern Web Guidance" skill for coding agents and enhanced DevTools debugging.

Chrome for Developers Blog

Summary

What: At Google I/O 2026, Chrome announced new features for web extension developers, including an AI-driven "Modern Web Guidance" skill for coding agents and expanded Chrome DevTools for agents with extension debugging capabilities. The update also brings granular team roles to the Developer Dashboard, private enterprise publishing for external organizations, and support for the browser global namespace.

Why it matters: Google is investing heavily in making extension development more accessible and powerful through AI, while simultaneously improving enterprise deployment and cross-browser compatibility, indicating a strategic push to invigorate the Chrome Web Store ecosystem and compete with other browser platforms.

Takeaway: If you develop Chrome extensions, explore the new Chrome DevTools for agents and the Modern Web Guidance extensions skill to streamline your AI-assisted development and debugging workflows.

Deep Dive

Google I/O 2026 highlighted that AI is making web extension development more approachable, evidenced by a doubling of monthly developer registrations and 17% of new extensions utilizing AI.
Google is expanding developer registration to over 120 additional countries to support the growing community.
A new "Modern Web Guidance" (MWG) extensions skill has been introduced for AI coding agents, providing web platform expertise, best practices, and modern API patterns for extension development.
Chrome DevTools for agents now includes support for extension debugging, allowing AI agents to programmatically install, uninstall, list, reload extensions, trigger actions, and inspect UI surfaces like popups and service workers.
The Chrome Web Store Developer Dashboard has new granular member roles (Admin, Item Manager, Viewer) to facilitate collaborative team management of extensions.
A new private enterprise publishing option allows developers to deploy extensions privately to approved external organizations directly from the Chrome Web Store.
Chrome now supports the browser global namespace, aligning with other major browsers to simplify the development and maintenance of cross-browser compatible extensions.
These updates aim to remove friction from the developer workflow and elevate the overall extension development experience on Chrome.

Decoder

Browser global namespace: A standardized global object (often browser or chrome) that provides access to WebExtensions APIs, making it easier to write extensions compatible across different browsers like Chrome, Firefox, and Edge.
Service worker: A type of web worker that runs in the background of a web page, separate from the main browser thread. They are commonly used in web extensions to handle events, intercept network requests, and manage caching.

Original Article

Original Article

Full article content is not available for inline reading.

Read the original article →

AI enterprisecloudgoogle

How Google plans to win the AI war

Google is aggressively embedding AI, like Gemini 3.5 Flash and "Ask YouTube," across its product ecosystem to safeguard its core businesses and maintain market leadership.

Axios

Summary

What: Google is integrating AI such as Gemini 3.5 Flash and features like YouTube's "Ask YouTube" across its products. The company is leveraging its extensive scale, distribution channels, and financial resources to deploy these AI capabilities.

Why it matters: This highlights Google's strategy to defend its dominant market position by infusing AI into its established offerings, demonstrating how legacy tech giants adapt to disruptive technologies by leveraging their existing infrastructure and user base.

Original Article

Google is aggressively integrating AI into its products to remain competitive while protecting its lucrative core businesses. The company leverages its scale, distribution, and financial resources to deploy AI like Gemini 3.5 Flash and features such as YouTube's "Ask YouTube." Despite challenges, Google aims to maintain market leadership by adapting its offerings swiftly without undermining its existing revenue streams.

Tech hardwareaerospace

SpaceX Postpones Launch of Newly Redesigned Starship

SpaceX's newly redesigned Starship launch was postponed due to a malfunctioning part in the launch tower, delaying a critical test of its re-entry tiles and Starlink deployment.

Wall Street Journal

Summary

What: SpaceX postponed the test flight of its upgraded Starship, designed to carry 100 metric tons to low-Earth orbit, after a launch tower component malfunctioned, though a re-attempt may occur today. The mission plans to deploy 20 Starlink simulators and evaluate the craft's re-entry tiles.

Why it matters: This delay highlights the immense engineering complexity and precision required for cutting-edge space launches, even for a seasoned company like SpaceX, underscoring the iterative and often challenging nature of hardware development.

Decoder

Low-Earth orbit (LEO): An orbit around Earth with an altitude between 160 km (99 mi) and 2,000 km (1,200 mi), commonly used for satellite internet constellations like Starlink and the International Space Station.

Original Article

Starship's launch has been delayed due to a malfunctioning part in the launch tower. SpaceX may try launching again today if the problem is fixed in time. The postponed test mission will deploy 20 Starlink simulators and stress-test tiles designed to protect the craft as it re-enters the Earth's atmosphere. The new investigation of Starship is designed to carry 100 metric tons to low-Earth orbit - the company's workhorse Falcon 9 vehicle can carry roughly 23 metric tons.

Tech aiwebadvertising

Google Pushes AI-Generated Ads Further Into Search Results

Google is expanding AI-generated advertisements within its search results, including brand promotions below AI Mode responses and discounted offers in standard search.

Wall Street Journal

Summary

What: Google began testing new AI-generated ad formats, placing brand and product advertisements below AI Mode responses and direct offers for discounted products in standard search results, but not yet in Gemini, with guardrails against AI hallucinations.

Why it matters: This move reflects Google's strategic imperative to integrate AI into its core advertising business, seeking to monetize AI Search while navigating user experience and the critical issue of AI reliability in commercial contexts.

Original Article

Google has started testing several new ad formats in both standard search and in AI Mode. AI-generated ads for brands and products will now appear below the responses to some prompts in AI Mode. Ads with direct offers for discounted products will be displayed in standard search results. The ads will not appear in Gemini for now. Google will protect the ads against AI hallucinations with strong guardrails.

Full article content is not available for inline reading.

Read the original article →

Tech researchhealth

Experimental Drug Yields Dramatic Weight Loss

Eli Lilly's experimental drug Retatrutide showed such dramatic weight loss in trials that some participants stopped taking it, with results comparable to gastric bypass surgery.

The New York Times

Summary

What: The experimental drug Retatrutide, developed by Eli Lilly, led to significant weight loss in human trials, with some participants discontinuing use due to losing too much weight, achieving results for the heaviest patients on par with gastric bypass surgery.

Why it matters: This signals the potential for a new generation of highly effective pharmaceutical interventions for obesity, which could dramatically alter public health approaches to chronic weight management, though side effects remain a concern.

Decoder

Gastric bypass surgery: A type of weight-loss surgery that involves making changes to your stomach and small intestine to change the way you absorb food.

Original Article

Retatrutide, an experimental shot by drugmaker Eli Lilly, was so powerful in human trials that participants stopped taking it because they thought they were losing too much weight. The results for some of the heaviest patients were on par with those seen with gastric bypass surgery. There are some side effects at higher doses, such as gastrointestinal side effects so unpleasant that some patients stop taking it. Eli Lilly has not yet applied for regulatory approval.

Digest devoured!

May 22

Home