AI Summary

Market Overview

• AGIX fell 1.16% DTD, underperforming the S&P 500 (-0.37%) and Dow Jones (+0.29%), but still significantly outperforming YTD with a gain of 29.38%.

• Among peer ETFs, semiconductor and AI-themed funds led: SMH rose 1.37%, IGPT gained 1.61%, and CHAT added 1.69%, while IGV declined 2%.

• Top positive contributors within AGIX holdings were Astera Labs (+5.42%), SK hynix (+5.61%), Micron (+6.82%), and TSMC (+4.15%), driven by AI demand and index inclusion events.

• Key stocks to watch include Micron (AI memory supply deal with Anthropic), Astera Labs (Nasdaq-100 addition), and Ciena (strong earnings & guidance), as these catalysts may sustain momentum.

News Highlights

1. Nvidia launched Halos for Robotics, the first safety framework for AI robots; Agility Robotics' Digit is already testing at Amazon, GXO, and Toyota, supporting Nvidia's hardware ecosystem and AGIX holdings like Nvidia.

2. Google DeepMind invested $75M in A24 for AI filmmaking tools; this partnership may accelerate AI adoption in creative industries, benefiting Alphabet and broader AI sentiment.

3. OpenAI expanded its Daybreak cybersecurity program with Patch the Planet and GPT-5.5-Cyber (85.6% benchmark score), highlighting AI's role in security and potential tailwinds for Palantir, CrowdStrike, and Zscaler.

4. HPE and Kamiwaza launched AI inference infrastructure with Nvidia, targeting production-grade deployment; this supports demand for Nvidia GPUs and HPE's AI servers.

5. SpaceX signed a $6.3B compute deal with Reflection AI to use Nvidia GB300 chips, underscoring sustained hyperscale AI compute demand.

RSS Highlights

1. OpenAI's lead is dwindling fast: market share declining, Microsoft withdrawing support, and annual losses growing eightfold, raising sustainability concerns.

2. The Silicon Valley Bubble (Part 2) analysis: OpenAI lost $21B in 2025 on $13B revenue, questioning the industry's focus on capitalizing consensus over genuine innovation.

3. Expert-aware quantisation achieved near-Q4 quality at near-Q2 size for MoE models, a breakthrough for efficient AI deployment on edge devices.

4. GLM-5.2, the most powerful open-weight LLM, now leads the AI Analysis Index with 75.3B parameters and a 1M context window, challenging proprietary models.

5. Cloudflare introduced temporary accounts for AI agents, allowing ephemeral Workers projects without registration, simplifying AI tool prototyping.

Performance & Benchmark

Attribution

Contribution of the three sectors within AGIX holdings today

Application: -0.33%

Hardware: 1.04%

Infrastructure: -0.89%

Top-performing stocks in AGIX holdings today

TOWER SEMICONDUCTOR LTD (10.5%)
Ciena Corporation (7.5%)
Vertiv Holdings Co (7.48%)
Micron Technology, Inc. (6.82%)
Renesas Electronics Corporation (6.49%)
NITTO BOSEKI CO LTD (6.05%)
SK hynix Inc. (5.61%)
Flex Ltd. (5.56%)
Astera Labs, Inc. (5.42%)
SANMINA CORP (4.95%)
GlobalFoundries Inc. (4.47%)
Teradyne, Inc. (4.36%)
Taiwan Semiconductor Manufacturing Company Limited (4.15%)
EATON CORP PLC (3.32%)
Texas Instruments Incorporated (2.92%)
Arista Networks Inc (2.88%)
Advanced Micro Devices, Inc. (2.65%)

Worst-performing stocks in AGIX holdings today

ServiceNow, Inc. (-2.14%)
TE CONNECTIVITY PLC (-2.27%)
TeraWulf Inc. (-2.31%)
Meta Platforms, Inc. (-2.32%)
Snowflake Inc. (-2.45%)
Cloudflare, Inc. (-2.54%)
Hut 8 Corp. (-2.73%)
APPLIED DIGITAL CORP (-2.98%)
Microsoft Corporation (-3.18%)
Broadcom Inc. (-4.67%)
Amazon.com, Inc. (-4.75%)
Alphabet Inc. (-4.99%)
Oracle Corporation (-5.0%)
CoreWeave, Inc. (-5.65%)
Netflix, Inc. (-5.82%)
Tempus AI, Inc. (-5.88%)
Palantir Technologies Inc. (-6.98%)
Arm Holdings plc (-7.22%)
Roblox Corporation (-8.27%)

Top five listed stocks contributing the most within AGIX holdings today

Arm Holdings plc (-7.22%)
Alphabet Inc. (-4.99%)
Amazon.com, Inc. (-4.75%)
Palantir Technologies Inc. (-6.98%)
Broadcom Inc. (-4.67%)

Observation

The strongest movers across peer ETF holdings today are listed below:

News

Funding & M&A

Ethan Thornton is trying to do everything all at once.

U.S. defense tech startup Mach Industries recently completed a $300 million Series C funding round, achieving a valuation of $1.8 billion and bringing total funding to $485 million. The company was founded by 19-year-old MIT dropout Ethan Thornton and is advancing six weapons programs, including vertical takeoff and landing aircraft, long-range anti-ship missiles, and air defense interception systems. To accelerate supply chain construction, Mach acquired Exquadrum, a solid rocket motor company with a 24-year history, for $150 million. Unlike the single-product strategies of Shield AI and Saronic, Mach adopts a multi-line parallel model and competes with Anduril. The founder stated that the U.S. struggles to surpass China in manufacturing capacity and must rely on innovation and rapid productization to gain an advantage. The company has secured 13 government contracts and aims to push three systems into mass production by the end of 2026.

Nearfield Instruments raises $380 million to accelerate AI chipmaking.

Dutch semiconductor equipment manufacturer Nearfield Instruments recently completed a total $380 million Series D funding round, led by Chongcheng Management, with participation from Winter, Singapore Investment Corporation, and the Qatar Investment Authority. This round values the company at $1.6 billion, making it the largest single unicorn in the Netherlands. Nearfield primarily provides advanced 3D metrology and process control systems to address the detection challenges of extreme-size structures in AI chip production, helping semiconductor manufacturers improve yield rates. The funding will be used to accelerate technological innovation, expand production capacity, and set up application research centers globally. As AI model scales continue to expand, market demand for high-precision advanced metrology equipment is growing rapidly, positioning Nearfield as a key player in the global chip supply chain.

Inference chip startup Groq raises $650 million to grow its cloud platform

Seven months after reaching a $20 billion chip licensing deal with Nvidia, Groq today announced a $650 million funding round. The round was led by Disruptive and Infinitum. Groq's LPU chip is focused on AI inference optimization, and its technology has been licensed by Nvidia and used to launch the Nvidia Grok LPU 3. The chip integrates 90 data channels, with bidirectional bandwidth reaching 2.5 Tbps, and supports 500 MB of high-speed SRAM, with automatic clock skew correction. Groq's cloud platform processes trillions of tokens weekly, serves 5 million developers, and has data centers across all 13 continents. The new funds will be used to expand capacity to 200 MW by 2027 and promote the deployment of LPX liquid-cooled servers. This move will strengthen Groq's competitive position in the AI inference cloud market.

AI chipmaker Groq confirms a $650 million raise and restaffs after Nvidia's $20 billion not-acqui-hire deal.

American AI chip startup Groq announced a new $650 million funding round after a non-traditional acquisition collaboration with Nvidia. Previously, Nvidia obtained Groq's LPU technology through a 1P licensing agreement and recruited its founder and core team. Groq was valued at $6.9 billion in its last round. After sharing its technical IP, Groq shifted its business focus to the 'New Cloud' service operated by the previously acquired company Definitive Intelligence. It now has 13 data centers in North America, Europe, the Middle East, and Asia Pacific, serving over 5 million developers. The new leadership team includes former Meta and xAI executive Alan Rice. Amid the rapid growth of the AI inference market, Groq faces competitive pressure from Nvidia but still has opportunities to break through in the cloud services sector.

AI networking provider Upscale AI raises $190 million at a $2 billion valuation.

Data center networking startup Upscale AI Inc. recently completed a $190 million extension of its funding round, bringing its total funding to $390 million and its valuation to $2 billion. The round was led by Premji Invest, with participation from Nvidia, Salesforce Ventures, Temasek, and others. The company focuses on developing scale-up and scale-out switches for AI workloads. Its self-developed SkyHammer chip enables deterministic latency and supports protocols such as UALink and optimized Ethernet. Additionally, its scale-out product line is based on Nvidia's Spectrum-X chip and integrates Microsoft's SONiC open-source operating system. The funding will be used to accelerate commercialization, and the company has already been evaluated by several hyperscale data center operators. As a key component of AI infrastructure, Upscale AI is building high-performance networking components through open standards.

Technical Breakthroughs

The AI world is getting ‘loopy’

At the Meta @Scale conference, Claude Code founder Boris Cherny emphasized that the agentic loop mechanism has become an important turning point in AI development. He pointed out that each step—from manually writing code to agents generating code, and now to multiple agents prompting each other in a continuous loop—is significant. This mechanism allows groups of agents to run in the background, automatically detect code architecture issues, merge duplicate abstractions, and submit pull requests without human intervention. This technology aligns with the concept of test-time compute, achieving incremental optimization through continuous token investment. Although running costs are high, it has great potential for complex tasks like code improvement. As model performance rapidly improves, the loop mechanism is expected to become a key technology for enhancing AI's practical productivity.

Nvidia introduces Halos for Robotics to bridge the physical AI safety gap.

Nvidia has launched Halos for Robotics, the industry's first robot safety system framework, integrating IGX Thor, Holoscan Sensor Bridge, and Halos OS core safety software to support building, testing, and managing full-stack AI robot applications. Agility Robotics is among the first adopters, with its humanoid robot Digit already being tested in Amazon, GXO, Schaeffler, and Toyota Canada factories. The Halos AI Systems Inspection Lab has received ANSI national certification and will assist partners in obtaining international safety certifications such as TÜV and UL. Facing the 2027 EU Machinery Regulation's compliance requirements for self-evolving AI robots, Nvidia is using platform-based safety standards to promote the large-scale deployment of physical AI, reduce human-robot collaboration risks, and accelerate the commercialization of Robotics as a Service.

Nvidia wants to cut data center water use, but that's not the same as fixing AI's water problem.

Nvidia recently introduced a warm water cooling system that claims to significantly reduce data center water consumption, even achieving 'near-zero water usage.' The system uses a closed-loop design where coolant is filled once and recirculated, with an inlet water temperature of 45°C and outlet of 55°C, dissipating heat through ambient air. It is expected to achieve 100% water savings on-site in suitable climates. However, Nvidia limits its accounting boundary to within the data center, ignoring the water footprint of electricity generation and chip manufacturing. Research shows that fossil fuel power generation consumes 1.17 to 2.2 liters of water per kilowatt-hour, accounting for two-thirds to three-quarters of a data center's total water consumption. The IEA predicts that natural gas and coal will still meet over 40% of new electricity demand by 2030. Therefore, while Nvidia's technology is effective on-site, the overall water stress of AI data centers will persist unless there is a shift to renewable energy.

Product Launches

Amazon is testing Alexa+ in India with Hindi support

Amazon is planning to expand its generative AI-powered conversational assistant Alexa+ to the Indian market and is inviting local users to test the Hindi version. The company has sent invitations to some customers via email, asking them to fill out a Hindi form by June 22 to join the beta program. The email notes that the test version may have bugs or voice recognition errors, but user feedback will be used to optimize the product. Alexa+ is not yet available in India; Amazon launched the English version of Alexa in India in 2017 and added Hindi support in 2019. This expansion aims to reach over 60 billion Hindi-speaking users. Alexa+ was first released in 2025 and expanded to countries including the UK and Canada in 2026, with free access for Prime users.

Industry Partnerships

SpaceX inks compute deal with Reflection AI, an open source AI lab

Open-source AI startup Reflection AI has reached a large-scale computing lease agreement with SpaceX, paying $1.5 million per month starting July 1, 2026, to use Nvidia GB300 AI chips at SpaceX's Colossus 2 data center near Mississippi, USA, until 2029. The total value of the agreement is up to $6.3 billion, and either party can terminate the contract with 90 days' notice after the first three months. This deal is smaller than SpaceX's partnerships with Anthropic and Google, but it is still one of the largest commitments in open-source AI infrastructure. Reflection AI sees this as core support for its open-weight model strategy, emphasizing the importance of open-weight models for ecosystem security after the US blocked Anthropic's closed-source models. The company was founded in 2024 by former Google DeepMind researchers, and this collaboration marks its first large-scale computing acquisition, providing ample resources for training extreme open-source models.

Google DeepMind bets $75M on AI’s future in Hollywood with A24 deal

Google DeepMind recently announced a $75 million investment in independent film studio A24 to jointly develop AI tools for filmmaking. This is Google DeepMind's first such collaboration with a Hollywood studio, aiming to create AI features that truly empower creators by working directly with professional artists. A24 is known for producing high-quality independent films, with recent projects involving Timothée Chalamet and Anne Hathaway. Google DeepMind CEO Demis Hassabis said that early collaboration with filmmakers helps develop tools that meet creative needs. Although the use of AI in the film industry remains controversial, A24 is not the first to explore it. Previously, Netflix acquired Ben Affleck's AI film tool company InterPositive, and Amazon's MGM studio has also established a department focused on AI development for film and television. This partnership may drive broader adoption of AI technology in the film industry.

The founder conference built for growth: TechCrunch Founder Summit pass rates increase June 26.

TechCrunch will host the Founder Summit 2026 in Boston on November 4, 2026, bringing together over 1,000 founders and investors, focusing on growth, execution, and building lasting connections. This summit is TechCrunch's flagship conference for founders, emphasizing peer learning, candid conversations, and actionable insights to help founders accelerate growth and directly connect with stage-matched investors. Early bird tickets are available until June 26 at 11:59 PM Pacific Time, with individual savings of up to $190 and group registrations of 4 or more enjoying up to 30% discount. Past topics have covered fundraising, product iteration, and team building, with speakers from top firms like Sequoia Capital, NFX, and Greylock. The 2026 agenda is being prepared, with more founders, operators, and investors to be announced. TechCrunch encourages attendees to submit roundtable or breakout session topics and participate in agenda voting. The summit will also offer exhibition opportunities for exhibitors to directly reach founders and decision-makers.

OpenAI expands Daybreak with Patch the Planet and the full GPT-5.5-Cyber release.

OpenAI Group PBC today fully upgraded its Daybreak cybersecurity program, launching the open-source patching initiative Patch the Planet, a new version of the Codex Security plugin, a partner program, and the official release of its strongest defense model, GPT-5.5-Cyber. OpenAI is collaborating with organizations such as Trail of Bits and HackerOne to fund expert researchers and use AI models to assist open-source project maintainers in fixing vulnerabilities. The first batch of participating projects includes over 30 critical open-source components such as cURL, Python, and Go. OpenAI emphasizes that the current security bottleneck has shifted from discovering vulnerabilities to efficiently patching them, with human review ensuring quality. The initial five-day sprint identified hundreds of issues and merged dozens of patches, while also discovering multiple exploitable vulnerabilities in browsers like Chrome, Safari, and Firefox. GPT-5.5-Cyber scored 85.6% on the CyberGym benchmark. The Daybreak Cyber Partner Program has attracted several security vendors including Accenture, Cisco, and CrowdStrike. OpenAI has also signed Trusted Access cooperation agreements with governments of the US, Australia, Canada, and the EU, continuing to promote industry collaboration and government regulation in AI security.

HPE and Kamiwaza rethink AI infrastructure for the inference era

As AI factories transform into future data centers, enterprise demand for production-grade AI infrastructure is increasing. Hewlett Packard Enterprise Co. (HPE) has launched the Unleash AI initiative, partnering with Kamiwaza Corp., Nvidia Corp., and others to build GPU-optimized inference architectures that address complex data processing and inference speed issues. The solution combines next-generation private cloud and storage platforms like HPE Alletra Storage MP X10000 to provide cloud-native, secure, and manageable database infrastructure, helping enterprises quickly move from experimentation to production. HPE executives say the new architecture can significantly improve performance and reduce costs without substantially increasing server and electricity expenses, promoting sustainable AI inference. This series of products and collaborations will help enterprises achieve efficient and scalable AI deployment.

Google forms a research partnership with A24 Films focused on AI filmmaking tools.

American tech giant Google is investing approximately $75 million in independent film studio A24 to explore the potential of artificial intelligence in the entertainment industry. This partnership marks Google's first direct investment in a film production company and will allow A24 to access its DeepMind AI research lab. The two parties plan to jointly develop new technologies to optimize filmmaking processes, including storyboard creation, rather than relying entirely on AI-generated content. A24 partners emphasize that the new tools will preserve creator control to avoid industry concerns. However, the deal has sparked controversy within Hollywood, with many directors and actors worried that AI could threaten traditional jobs and alter the original intent of works. Although A24 has accepted controversial funding in the past, this collaboration with a tech giant is seen as a significant signal of industry change.

产业合作

Intrinsic unveils next-gen accessible modular automated industrial AI robotic assembly.

Google's Intrinsic unveiled an AI robot prototype at the Automate 2026 exhibition, introducing a modular 'smart cell' reference design based on the IntrinsicOS platform for factory automation upgrades. The solution uses the Flowstate simulation environment and reusable skill modules to simplify traditional hundreds of hours of programming into drag-and-drop operations, significantly lowering integration barriers. Intrinsic has partnered with Foxconn, FANUC, Trinity Automation, and MartinSystems to jointly promote the rapid deployment of AI robots from simulation to production lines. The platform emphasizes hardware interoperability and supports CNC system integration, helping manufacturing enterprises of all sizes implement physical AI applications, marking a new phase of software-defined industrial automation.

RSS Subscription

Pluralistic: The (real) dead economy theory (17 Jun 2026)

Recently discussed 'dead economic theories', pointing out the failure of financial markets in accurately assessing assets, especially with the rise of AI technology, as capital continuously flows into ineffective investments and bubbles. Taking Elon Musk as an example, most of his business projects since 2020 have not succeeded, instead leading to a reduction in funding for actual productive activities, affecting important areas such as cancer research.

You Got Faster. Your Company Didn’t.

Due to the application of artificial intelligence, the personal work efficiency of team members has significantly improved, but the overall company operation may not necessarily speed up. When engineers write technical documentation, although they save time, the AI-generated documents require more reviewers to verify, causing the overall process to slow down. Therefore, while individual work speed increases, team collaboration efficiency decreases, reflecting a need to strengthen responsibility toward readers.

Logic for Programmers v0.15, Livecoding

Logic for Programmers v0.15 is released, the first release candidate. All content has been edited and proofread. The next version will be 1.0 and will be available in print. Meanwhile, the author is testing a small PDF format suitable for phones and computers. Recently attended the first meetup of livecode.chicago, showcasing two languages for live music coding: Strudel and CLAVIER-36. Interested readers are encouraged to learn about these languages.

Formalizing a ring theorem with Lean 4 and Claude

I tested Claude's ability to generate Lean 4 code for proving theorems, specifically a theorem about partial fraction decomposition. After 11 iterations, Claude finally generated a proof containing five 'sorry' sections representing unproven assumptions. Although the generated code failed to build successfully, Claude's proof strategy and approach were very helpful for manually proving the theorem.

Windows stack limit checking retrospective, follow-up

Aaron Giles mentioned that when porting Windows to ARM32 and AArch64, there is a detail about arm64 stack limit checking that was not mentioned: why the parameter uses the x15 register instead of the usual x0 register. Since the stack probe function needs to handle incoming parameters, using a custom calling convention to employ a temporary register not used for receiving parameters is the solution. This method is applied in multiple processor architectures, for example, PowerPC uses r11, AArch64 uses r16 and r17.

How a Microsoft product from June 1979 led to the IBM PC

In June 1979, Microsoft's 8080 Basic installations exceeded 200,000 for the first time, and on June 18, it released a new version of Basic for the Intel 8086. IBM wanted to collaborate with Microsoft to obtain Basic, an operating system, and industry advice to support the development of its new desktop computer. Ultimately, Microsoft's 8086 Basic became a key selling point for the IBM PC and was pre-installed on all PC and PS/2 products, even though these products were not equipped with tape drives.

Flax debugging: making a hash of things

While debugging a JAX/Flax NNX training loop, I noticed that the model's loss remained stuck at 10.82, indicating that the model might be guessing randomly. By printing the gradients, I confirmed they were reasonable, but I suspected that the optimizer's parameter updates were not effective. Eventually, I discovered that Flax's NNX API uses in-place modification for parameter updates, which may have prevented the updates from executing correctly.

OpenAI’s lead is dwindling fast

OpenAI's market leadership is rapidly weakening, mainly due to a lack of competitive barriers and a continuous decline in market share. Microsoft, its largest investor, has begun to gradually withdraw support and is even considering outsourcing its main products to China, which is seen as a strong negative signal. Additionally, OpenAI's cash burn rate is exceeding expectations, with annual losses growing eightfold, making it unsustainable.

Debugging on Prod

Errors that occur in production environments are often difficult to debug. The author encountered a 500 error while updating the website design and code. After multiple investigations, the problem was traced to an invalid HTTP header in the caching mechanism, which caused an error in the PHP-FPM environment but not in the local environment. Ultimately, the issue was resolved by simplifying the code logic.

10Gb/s Ethernet: switching to a Broadcom SFP+ module

After upgrading to 10Gb/s Ethernet at home, I found that the MikroTik S+RJ10 module I was using ran too hot, causing network instability. To solve this, I replaced it with a 10Gtek ASF-10G-T80-INT module, which uses a Broadcom chip and offers lower power consumption and longer transmission distance. After the replacement, the network returned to normal and the temperature issue improved.

Partial fraction decomposition

Partial fraction decomposition is commonly used to compute integrals, but its concept can be more broadly applied to the factorization of polynomials. Any polynomial with real coefficients can be decomposed into a product of linear and irreducible quadratic factors; however, for high-degree polynomials, computing this decomposition is difficult. For the ring of integers, any rational number can be expressed as a sum of fractions whose denominators are powers of primes, with numerators coprime to the denominators.

Ada Palmer – Machiavelli is the most misunderstood thinker of all time

This episode invited historian and science fiction writer Ada Palmer to discuss Niccolò Machiavelli, one of the most misunderstood thinkers in history. Machiavelli was a senior diplomat in Florence who was fired, tortured, and exiled after being mistakenly suspected by the Medici family of involvement in a coup. During his exile, he wrote "The Prince" and dedicated it to Lorenzo de' Medici, who had tortured him, in an attempt to offer political wisdom to Florence. Palmer also mentioned that her new podcast will delve into Machiavelli's ideas, with the first season focusing on this topic.

Accenture: Then and now, and how it may signify things to come

Accenture's stock fell about 18% after a disappointing quarter, dropping more than 50% from its 52-week high. Although the company had hoped for a transformation driven by artificial intelligence, current performance shows that generative AI has not brought significant returns, a view supported by findings from multiple research institutions.

‘Popa’ Botnet Linked to Publicly-Traded Israeli Firm

Over the past four years, an Android botnet named Popa has forced millions of consumer TV boxes to relay internet traffic related to ad fraud, account takeovers, and large-scale data scraping. Recently, researchers from several security companies found that the Popa botnet is linked to NetNut, a "residential proxy" provider operated by Israeli publicly traded company Alarum Technologies Ltd.

Stealth Startup Spy #349

A green energy infrastructure startup exited stealth mode with a $27 million seed round, while a former McKinsey consultant is building an accountability layer for ROI on enterprise AI agents, and another researcher is focused on discovering microbial dark matter.

Pluralistic: AI digital sovereignty risk doesn't exist (18 Jun 2026)

The current risk of digital sovereignty does not exist. As shown by 'problem + blockchain = problem - blockchain', the complexity of blockchain does not solve the trust problem but may instead exacerbate it. Global dependence on US tech platforms makes digital sovereignty a real issue, but if countries can abolish laws that protect US tech giants, it could turn into an opportunity for digital sovereignty.

Why doesn't GetLastInputInfo() return info for the user I'm impersonating?

The customer wants to obtain the last input time of all logged-in users through a Windows NT service process. They plan to use WTSQueryUserToken() to get the user token and perform user impersonation, but they found that GetLastInputInfo() always returns the input information of the service session, and the service session is not interactive, so it cannot obtain user input information from other sessions. GetLastInputInfo() only provides input information for the session that calls the function. To obtain information from other sessions, an auxiliary process needs to be started in the target session to collect data and send it to the service.

Book Review: The Great When by Alan Moore ★★★★☆

This book is the most verbose work I have ever read. Author Alan Moore uses ornate language to turn an ordinary story into a precious narrative. Although the plot structure is relatively simple, the story is still engaging. The characters in the book are similar to those in the books I have read about wartime Black Britain, which feels familiar. Although I have reservations about the theme of 'the decline of magic in England', the overall reading experience is enjoyable, and I look forward to the release of subsequent works.

Windows ME released June 19, 2000

Windows Millennium Edition (Windows ME) was released on June 19, 2000, as the successor to Windows 98 SE and the least successful operating system from Microsoft since the 1980s. Although Windows ME improved disk defragmentation and plug-and-play, its stability and compatibility issues caused widespread user dissatisfaction, and many chose to continue using Windows 98 SE or Windows 2000. The release of Windows ME is considered a major misstep by Microsoft, marking a decline in its operating system reputation.

GLM 5.2 playing text adventures

The performance of the new GLM 5.2 model in text adventure games has attracted attention. Its output token cost is $4.4 per million, slightly higher than Google's $3. Through comparative testing, GLM 5.2 achieved 15% fewer achievements in the game than Gemini 3 Flash. This difference is statistically significant, indicating slightly inferior performance, but it still outperforms Gemini 2.5 Flash.

Glassblowing #2: Making a tungsten lamp and (bad) vacuum diode

I made a tungsten bulb and experimented with a vacuum diode. The tungsten filament glows at 2500°C, but since it burns in contact with air, I sealed the bulb in a vacuum glass. Eventually, the bulb emitted bright white light at a current of 200 to 400 mA, despite issues with heat and short lifespan. Additionally, my diode experiment only conducted 1 microampere at a bias of 700 V, mainly due to gas in the tungsten filament and the anode area being too small.

Understanding the rationale behind a rule when trying to circumvent it

When implementing callback functions related to processes and threads, best practices emphasize that callbacks must execute quickly and must not block, including avoiding user-mode service calls, registry calls, and synchronization with other threads. Callback functions should use system worker threads to handle time-consuming tasks to ensure system performance is not affected. Documentation updates indicate that when using system worker threads, never wait for tasks to complete, so as not to violate the intent of asynchronous processing.

The 100,000 whys of AI

The difference between AI-generated text and human writing has sparked widespread debate, especially since the output of large language models (LLMs) is almost indistinguishable by statistical tests. Recent research shows that many book covers for '100,000 Whys' on Amazon are highly similar, reflecting a characteristic of AI-generated content: the same prompt can lead to 80% identical output. This phenomenon is evident not only in book titles and cover designs but also in author names showing obvious repetition, indicating that LLMs use similar complex styles when responding to routine prompts.

Pledging Another $400,000 to the Zig Software Foundation

Mitchell Hashimoto announced that he will donate an additional $400,000 to the Zig Software Foundation, bringing his total donations to $700,000. He expressed appreciation for the technical progress and community management of the Zig project, and emphasized the project's uniqueness and pursuit of quality in the open-source field. Hashimoto also mentioned that although his views on AI are not entirely aligned with the Zig Software Foundation's policies, he still respects the foundation's philosophy and efforts.

On Vulgar Materialism

In 1914, although people believed that a European war would lead to economic disaster, the war eventually broke out, proving that view wrong. Similarly, in 2022, people thought that Russia's invasion of Ukraine would cause economic losses for Putin and his oligarchs, but Putin still launched the war, and the oligarchs' assets were seized. This shows that the materialistic perspective of modern society is overly simplistic, ignoring the complexity of political power and historical context.

All pieces on a 6 by 5 board

Using Z3/Python code generated by Claude, I solved a chess puzzle: on a 6×5 board, place a king, queen, two bishops, two knights, and two rooks, with the two bishops on opposite-colored squares and no piece attacking another. The code found 192 solutions, 24 of which are unique.

Pluralistic: How the Epstein Class recruits (20 Jun 2026)

Recently, a data leak from Peter Thiel's secret society Dialog has raised concerns about members rating each other and discussing topics like "race science." This incident reveals how some wealthy individuals maintain their social circles by charging others, highlighting that creative workers should be wary of potential scams disguised as "help."

Reading List 06/20/26

The U.S. Congress reached an agreement on a new housing bill that removes cumbersome rental construction provisions. Meanwhile, General Motors is entering the large-scale battery market, and semiconductor startup Phoenix Semiconductor is producing alternative microchips for military equipment. Additionally, China controls 70% of the global indium phosphide supply and will impose export restrictions in 2025, potentially affecting global AI data center construction.

Which Copyleft License is Suitable for an SVG?

Scalable Vector Graphics (SVG) is an XML-based image definition format that maintains clarity at any size. Although SVG can be viewed as an image, it is actually code that generates the image, so its software nature should be considered when choosing a license. While Creative Commons licenses are suitable for creative works, they are not appropriate for software; it is recommended to use software-specific licenses like the GNU General Public License (GPL). In summary, SVG is more akin to software than to traditional images.

I know Kung-fu

In The Matrix, Neo's ability to gain skills by uploading information prompts reflection on knowledge acquisition. Although modern technologies like ChatGPT can provide abundant information, true knowledge still requires time and understanding to accumulate. Simple information downloads cannot replace deep learning and thinking; the formation of knowledge depends on the digestion and internalization of information.

Glassblowing #3: A better thermionic diode

In my latest vacuum tube experiment, although technically feasible, the excessive glass between the anode and cathode prevented effective electron flow. By enclosing the filament within the anode, the accumulated negative charge was shielded, enabling continuous current flow. After a series of adjustments and tests, the filament current was 286 milliamps at 4 volts, power was 1.1 watts, and reverse leakage current was 22 nanoamps at 200 volts, which performed quite well. After 24 hours, the device's performance remained the same as the initial test, showing it could operate normally without a pump connection.

Lobachevsky's integral formula

The theorem proposed by Lobachevsky applies to even functions f with period π, and is significant in Fourier analysis and signal processing, especially in the special case where f(x) = 1.

Consistency, but in excellence, not appearance

In visual design, excessive pursuit of consistency may suppress individual excellence. Taking macOS icons as an example, although the Creator Studio icon is visually consistent, it lacks uniqueness and iconic character, while earlier icons each had distinctive shapes and materials, showcasing individual excellence. Pursuing individual excellence rather than mere visual consistency can give design deeper meaning and purpose.

Pluralistic: Good politics (22 Jun 2026)

Brazilian President Lula and his Workers' Party recently achieved a major political victory by legislating a five-day workweek, aiming to improve the living conditions of the working class. Although Brazil's working poor are still bound by a six-day workweek, the new policy has been widely welcomed, even forcing opposition party members to vote in favor. In contrast, UK Labour leader Keir Starmer, facing a huge majority, failed to improve people's lives and instead adopted unpopular policies such as spending cuts, demonstrating the consequences of bad politics.

Stealth Startup Spy #350

Stealth startup updates: Former Beamery CEO goes stealth, Caltech machine learning researcher develops AI for lab protocols, former SpaceX engineer works on nuclear fuel conversion technology. Currently, our 'Company Launch Tracker' tracks multiple founders and emerging startups, providing real-time updates on stealth activities to ensure you get the latest information first.

In memory of the man who put red and green squiggles under words

Tony Kruger is remembered for his contributions to multiple versions of Word, especially the red and green wavy line spell-check feature he first introduced, which greatly improved user experience. He also ported the game Chip’s Challenge to Windows through reverse engineering without source code. Today, nearly all word processors use similar spell-check features, and Tony's work has had a profound impact.

Cybersecurity for the paranoid business traveler

In business travel, employees face threats such as espionage and extortion, so taking appropriate security measures is crucial. It is recommended to avoid using USB flash drives, use personal charging devices whenever possible, reduce the number of mobile apps, and use VPNs cautiously to lower the risk of surveillance. Additionally, carry essential passwords and files, and use a cheap laptop for travel tasks to protect personal information.

AMD Athlon: AMD’s game-changing CPU from 1999

On June 23, 1999, AMD released the highly anticipated Athlon CPU as the successor to its successful K6, and it officially launched on August 9. The Athlon was the first non-Intel x86 CPU to surpass Intel's fastest CPU at the time, helping AMD stand out in the competition and keep pace with Intel in clock speed from 1999 to 2002, ultimately achieving profitability. Later, the Athlon 64 extended the x86 architecture from 32-bit to 64-bit, further boosting AMD's market competitiveness.

Everything you say CAN and WILL be used against you

In the digital age, users often unknowingly give up their rights when registering for services, especially privacy and the right to silence. Recently, Anthropic announced support for customer identity verification, requiring users to upload identification to access services, meaning their identity is permanently tied to all actions, which could lead to unexpected consequences, such as being considered a suspect in a security audit. By agreeing to service terms, users are effectively waiving their right against self-incrimination, and all behavior and information may be used as adverse evidence.

Happy Father's Day.

As the father of twin boys, I often wonder when I became a responsible family man, which leaves me feeling confused. Ten years ago, I wrote a short story about my father, trying to describe him through clocks because he loved clocks so much. I hope you enjoy this 'Ode to My Father.' Happy Father's Day to everyone.

Queens on a prime order board

The n-queens problem is to place n queens on an n×n chessboard so that they do not attack each other. When n is a prime number greater than or equal to 5, queens can be placed on lines with slopes from 2 to n-2. All solutions for n=5 have this form, and for larger prime n, some solutions also follow this pattern.

Expert-aware quantisation: near-Q4 quality at near-Q2 size?

Research shows that by applying expert-aware quantization to a model, quality close to Q4 can be achieved at a size near Q2. Specifically, for the Qwen3.6 35B-A3B model targeting C++ programming tasks, using a selective quantization strategy that keeps 64 "hot" experts at high precision while reducing the other 192 "cold" experts to 2 bits, a reading score of 1.635 was achieved in a 14GB model, recovering nearly 90% of the quality gap between Q2 and Q4. This approach demonstrates the importance of expert selection during quantization for performance.

Pluralistic: The Big Con (19 Jun 2026)

Currently, pyramid schemes in the United States (such as Amway) have not only affected the financial situations of countless participants but also deeply influenced modern politics. Many self-proclaimed successful entrepreneurs, like Trump and Musk, actually rely on government bailouts and subsidies, masking their true financial conditions. People involved in these schemes often mistakenly believe they are the only losers, failing to realize the fraudulent nature of the entire system.

Full Page Paralysis

"Blank page paralysis" usually refers to the difficulty of starting a task, but in reality, finishing a task is the true challenge. In software and logistics, this is known as the "last 90%" and the "last mile," because the final stage is often harder. Completion makes the work real and open to judgment, while an unfinished state allows for the illusion of potential perfection.

Premium: The Silicon Valley Bubble (Part 2)

Recently, OpenAI released its audited financial reports for 2024 and 2025, showing a loss of $21 billion in 2025 on revenue of $13.07 billion, with about 6.6% of revenue coming from SoftBank. Although the company attempts to obscure its financial situation through complex accounting methods, the stark imbalance between its massive spending and revenue raises questions about the Silicon Valley bubble, suggesting the industry is no longer focused on genuine innovation but is instead沉迷于 capitalizing on consensus and theoretical futures.

The data black hole at the center of AI

One definition of intelligence is sample efficiency—how much data is needed to operate fluently and effectively in a specific domain. In recent years, although data distribution has improved significantly, progress in training sample efficiency has not been obvious. Frontier models are trained on tens of trillions of tokens, far exceeding the roughly 200 million tokens a human encounters in a lifetime, making AI far less efficient than humans at learning complex tasks.

The Goldilocks Principle in Fantasy Strategy

While writing history, I spent a huge amount of time on Heroes of Might and Magic II, more than on any other game. Although I also played Heroes of Might and Magic III, due to the special circumstances during the pandemic, my obsession with Heroes of Might and Magic II was deeper. Through this experience, I began to think about why I am so drawn to fantasy strategy games and decided to explore other similar games from the 1990s, such as Warlords III, Age of Wonders, and Disciples: Sacred Lands.

What does it mean when the bottom bit of my HMODULE is set?

The value of an HMODULE is usually the base address of the DLL or EXE it represents. These base addresses are always multiples of 64KB, so the bottom 16 bits are typically zero. However, if the bottom bit is set to 1, it indicates that the DLL was loaded as a data file rather than in the normal way. In this case, the DLL does not appear in the loaded module list, and its dependencies and initialization are not handled—it is just some bytes mapped into memory.

Converting Coal Plants to Natural Gas

Since 2008, the transformation trend of U.S. coal power plants has been significant, with over 140 coal plants converted to use natural gas, driven mainly by strict environmental regulations and the shale gas boom. Coal's share of U.S. electricity production has fallen from 57% in the mid-1980s to about 16% today, and this trend is expected to continue. With stricter environmental standards and lower natural gas prices, many coal plants have chosen to switch to natural gas to reduce emissions and operating costs.

Jay Miner, Atari and Amiga computer designer

Jay Miner was a pioneer in video games and computing, designing the Atari 2600 game console and the Amiga computer. He served during the Korean War, later earned a degree in electrical engineering, and worked at companies such as Atari and Compaq, driving many technological innovations. Although his contributions to medical devices were significant, he died in 1994 at the age of 62 due to health issues.

There Are No Instances in atproto

AT Protocol has no instances. The centralization problem of traditional social media is that users rely on a single platform, while AT Protocol separates hosting from aggregation, allowing users to freely choose hosting services and applications. Unlike Mastodon's concept of instances, AT Protocol provides a more flexible decentralized structure, enabling users to easily switch hosting services and try new apps.

Arp 29: The fireworks galaxy

Arp 29, also known as the "Fireworks Galaxy," is named for its bright star formation and the ten observed supernovae. The eastern arm of this galaxy is noticeably brighter and denser than the rest, and it is located between 10 million and 30 million light-years away. Due to interstellar dust obscuration, its surface brightness is low, making photometric measurements unreliable.

Pluralistic: Google's new remote attestation scheme is every bit as terrible as its old remote attestation scheme (12 Jun 2026)

Google's new remote attestation scheme is just as bad as the old one, revealing its unethical and deceptive behavior in the market. Google has lost three federal antitrust cases for anti-competitive conduct, with losses exceeding $20 billion, and its business strategies have further solidified its market dominance by manipulating the advertising market and restricting competitors. Although Google started as an open web ecosystem, its subsequent actions have gradually distorted that original intent.

Premium: The Silicon Valley Bubble (Part 1)

OpenAI and Anthropic have filed for public listings and begun competing for funding, but both companies are losing billions of dollars annually and lack a path to profitability. OpenAI is expected to go public within the next year, but its financial situation is concerning, requiring $865 billion to be raised within four years to meet its commitments. Meanwhile, Anthropic is pursuing a $35 billion private credit deal to support its growth plans, though its financial transparency is questioned.

I can never fully embrace LLMs for code

My sister earned a computer science degree a few years ago, and I have been encouraging her to understand the world of programming. Although I once told her she didn't need to understand code line by line, now I find myself needing to understand every part when using large language models to generate code, because I cannot submit code I don't understand. While others trust these tools and generate code quickly, I find that this need for understanding prevents me from fully leveraging the speed of code generation and becoming an efficient engineer.

Gadget Review: TP Link EH210 Ethernet Splitter (USB-C) ★★★★★

When setting up Ethernet at home, I laid CAT6 cables in every room. Although some rooms have many devices, the gym only has a connected TV. Recently, I bought the EH210 USB-C powered Ethernet splitter, which is easy to connect, wakes up quickly, supports video streaming, performs well, and is reasonably priced—suitable for users who don't need a complex switch.

Intel’s Pentium FDIV bug and recall

In June 1994, a math professor discovered that Intel's new Pentium CPU had an FDIV bug, causing errors in floating-point division. Intel recalled the 60 and 66 MHz Pentium CPUs, losing up to $475 million and suffering a major public relations blow. Although the bug affected few users, Intel's initial response angered industry experts, ultimately prompting the company to improve its validation process for future CPUs.

I Am Not a Reverse Centaur

In a blog post, the author expressed concerns about using large language models (LLMs) to generate code. Although he noticed an increase in contributions to open-source projects, most came from machine-generated code. He decided to no longer accept pull requests without prior discussion, believing such contributions lack genuine human involvement and may affect project quality. He feels uncertain about the future of open source and worries about the trend of machines replacing human programming.

You can finally power on a Mac remotely

Apple has finally allowed users to remotely power on a Mac without pressing the power button. This feature is available in macOS 26.5 and supports Mac mini, Mac Studio, and iMac models released in 2024 and later. Users can configure their Mac to automatically start when power is restored via System Settings, greatly facilitating use in labs and with portable devices.

Apple Is Going to Raise Device Prices — but When?

Apple CEO Tim Cook warned in an interview with The Wall Street Journal that Apple will raise product prices due to significant increases in RAM and SSD costs. Analyst Mark Gurman predicts that price increases may coincide with the upcoming back-to-school promotion, especially around the launch of new iPhones in September. Although Apple rarely raises prices during a product cycle, the global RAM shortage may force this move.

Gurman says the second-generation iPhone Air, coming in early 2027, will feature a 0.5× ultra-wide second camera.

Apple plans to launch the second-generation iPhone Air in the spring of 2027, aiming to enhance the device's appeal. The new model (codenamed V62) will add an ultra-wide-angle rear dual camera and has entered advanced testing stages.

Criterion Collection: The Complete Kubrick

This 30-disc set includes 4K restorations of Kubrick's 13 feature films and 3 short films, along with their original soundtracks and 5.1 mixes, over 25 hours of interviews, documentaries, and behind-the-scenes material, and a new 4K restoration of Vivian Kubrick's documentary 'Making The Shining'. It also includes new commentary tracks recorded by filmmaker Lee Unkrich and author Michael Benson.

What was Matt thinking?

In 1995, high school student Matt Wright created Matt's Script Archive, sharing simple website tools such as contact forms and guestbooks, with WWWboard becoming a widely used web forum application. However, these scripts had serious security issues, such as the CVE-1999-1479 vulnerability, which allowed attackers to execute code as root on the server. Although Matt's scripts met user needs at the time, poor design and lack of updates led to security risks, ultimately prompting other developers to create alternatives to address these problems.

Dickover of the Week: The Observer

Recently, users browsing The Observer website discovered that the site shares users' personal data with 161 partners, sparking strong concerns about privacy protection. Users expressed anger over this large-scale data collection, calling it unacceptable and even comparing it to historical violations of personal privacy.

sqlite-utils 4.0rc1 adds migrations and nested transactions

sqlite-utils 4.0rc1 was released on June 21, 2026, adding database migration and nested transaction features. This version includes support for the sqlite-migrate package, allowing users to easily manage database migrations, and introduces db.atomic() transaction management, simplifying the use of nested transactions. This version also includes some backward-incompatible changes, such as modifications to the default floating-point column type and the removal of support for Python 3.8.

sqlite-utils 4.0rc1

sqlite-utils version 4.0rc1 has been released, adding migration and nested transaction features. For more details, please visit the relevant link.

Before and After: MacOS 27 Golden Gate Beta 1’s App Icons

During WWDC, Basic Apple Guy noticed an update to the icon design of the macOS Golden Gate beta, with more vibrant colors and several icons adjusted. Notably, the Journal icon's liquid glass effect changed significantly. Although the flat design of the icons remains controversial, the overall changes are seen as positive, indicating that the design trend is moving in the right direction.

Temporary Cloudflare Accounts for AI agents

Cloudflare has introduced a temporary account feature, allowing users to create Cloudflare Workers projects with the command "npx wrangler deploy --temporary" without needing to register an account. The project will remain online for 60 minutes, during which users can obtain the project's URL and choose to convert it to a permanent account.

Bobby Prince has died

The music of Doom is not only highly acclaimed in the gaming world but also sounds great on various sound cards. Although many consider "At Doom's Gate" to be the most iconic track, the author tends to think that tracks like "I Sawed The Demons" and "Evil Incarnate" have more depth and power. The PlayStation version of Doom used other composers because the original music creators were not paid royalties; while the music quality is decent, it does not match the original's standard.

The doom justifies the valuation

During my two-week trip to Berkeley, I observed a kind of frenzy about AI doomsday, a phenomenon that reflects people's justifications for their life choices. Despite technological progress, current technology is not sufficient to support its valuation. Media and public panic about AI are being amplified, leading to unfounded promises about the future. As doubts grow about the claim that "AI will replace all jobs," society needs to reflect on how to build a sustainable economy and society.

This Week in Package Management: 20 June 2026

In the fifth week's update, sbt 2.0.0 migrates to Scala 3, requires JDK 17, and introduces a new caching layer; pnpm 11.7 adds the frozenStore setting, supporting read-only operations; npm 12.0.0-pre.1 is the first prerelease enabling v12 default settings, blocking installation scripts for dependencies not listed in allowScripts. Other updates include Docker Engine 29.6.0's new image attestation endpoint and Homebrew 6.0.2's security enhancements.

Another One for the ‘Sorry, We Used to Be Crap’ Truth-in-Advertising File: Carlsberg Beer

Danish brewery Carlsberg admits that its long-promoted slogan 'probably the best beer in the world' is not true, and due to declining sales and negative consumer feedback on its beer's taste, it has introduced a new recipe and a more honest marketing strategy. The new ad campaign claims 'probably not the best beer in the world' and has renamed its flagship product 'Danish Pilsner', with an alcohol content of 3.8% and improved flavor.

‘What’s the Deal With Old Guys and Giant Glasses?’

Early adoption of new technology is often seen as a privilege of the young, but Snap Specs may change that perception, even being sold directly in nursing homes.

Quoting Sean Lynch

The real advantage of MCP over skills and command-line interfaces is its ability to isolate the authentication process outside the agent's context window, or even completely away from the agent. Ideally, MCP can serve solely as an authentication gateway for APIs, which is still a significant improvement.

Trump Mobile T1 Phone Is a Gold-Painted Two-Year-Old HTC U24 Pro

The Trump Phone T1 is touted as 'Made in America', but technical analysis shows it is nearly identical to the two-year-old HTC U24 Pro, which was produced by Taiwanese company HTC using Chinese parts. The Trump Phone sells for $500, slightly higher than the HTC brand's $470.

Fox to Buy Roku Streaming Service in $25 Billion Deal

Fox Corporation announced the acquisition of streaming platform Roku for approximately $25 billion, the largest deal in Fox's history, aimed at strengthening its ad-supported streaming business. The acquisition will combine Fox's streaming services with Roku's ad-supported channels to better compete with rivals like Amazon and Netflix for advertising revenue.

Snap Launches Ad Campaign for Specs Starring Michael Caine

This passage discusses the fashionability of a certain type of glasses and their impact on the wearer, suggesting that these glasses are associated with power, and mentions that the wearer may experience discomfort in the ears.

Jerry Seinfeld Tries Out Snap’s Specs

"Hey, buddy, nice glasses." Seinfeld's father also tried on these glasses.

Domino’s Admitted Their Pizza Tasted Like Cardboard

In 2021, Jeff Haden reported on Inc. about Domino's Pizza's successful turnaround, noting that the company had admitted its pizza tasted like cardboard. This move is similar to Verizon's business approach, demonstrating how companies respond to criticism.

Verizon, Formerly Menace Mobile

Verizon launched a new ad set against the backdrop of Austin Powers, in which Dr. Evil (played by Mike Myers) proposes launching a "Menace Mobile" wireless service, only to be rebutted by his son Scott, who says it's just typical phone company behavior. Scott emphasizes that Verizon has eliminated activation and upgrade fees and is changing its old business model. The ad has drawn attention for its candid acknowledgment of the company's past fee practices.

Datasette Apps: Host custom HTML applications inside Datasette

Datasette has launched a new plugin called datasette-apps, which allows users to host custom HTML+JavaScript applications within their apps. These applications run in a restricted iframe sandbox, can execute read-only SQL queries, and support write operations after configuration. The plugin aims to combine Datasette's flexible backend with self-contained HTML frontends to provide powerful functionality. Users can experience it through a GitHub login demo instance.

SpaceX, newly public, to acquire Cursor for $60 billion in SpaceX funny-money stock

SpaceX announced the acquisition of AI startup Cursor for $60 billion in stock, a deal revealed shortly after the rocket manufacturer's recent Nasdaq listing. Founded in 2022, Cursor has achieved rapid growth thanks to its popular AI coding tools, with annual revenue exceeding $10 billion in 2022. The acquisition is expected to close in the third quarter of this year, pending regulatory approvals.

Cotypist – Smart Autocomplete Utility for Mac

Cotypist is an intelligent writing tool that runs locally on Mac, quickly predicting the user's next word to boost writing efficiency. It supports multiple Mac applications, requires no registration or credit card, offers a 30-day professional trial, and can be used for free afterward. By learning the user's writing style, Cotypist helps users express ideas more fluently while protecting privacy by avoiding data upload to the cloud.

datasette-acl 0.6a0

Version 0.6a0 of datasette-acl has been released, expanding permission management from a table-only permission system to a general resource sharing system. This update, primarily led by Alex Garcia, aims to provide finer-grained resource access control for multi-user Datasette instances.

Tim Cook, in an interview with the WSJ: 'Unfortunately, price increases are unavoidable'

Apple CEO Tim Cook stated in a Wall Street Journal interview that due to soaring memory and storage chip costs, Apple plans to raise product prices to address this challenge. He noted that price increases are inevitable, despite the company's efforts to mitigate cost pressures and protect consumers. Apple is expected to release the iPhone 18 series, including a new foldable iPhone, in September.

New Domain for Sign In With Apple and iCloud+ Hide My Email

Apple will unify the email domains used for "Sign in with Apple" and "iCloud+ Hide My Email" this summer, with the new shared domain being private.icloud.com. Future addresses will use this domain, while existing old domain addresses will continue to work and forward emails. This change may cause some services to block the domain, potentially affecting the effectiveness of "Hide My Email."

NetNewsWire Status

Over the past year, after retiring, I focused on modernizing NetNewsWire, completing 2,188 commits. The main work included adopting Swift structured concurrency, optimizing the UI, fixing numerous bugs, reducing resource consumption, and enhancing performance. Additionally, we added support for the Cache-Control header, optimized iCloud sync, and introduced diagnostic and error reporting features. Although the foundational work is not yet complete, the development experience of the app has greatly improved. Thanks to all contributors for their support.

Show your hands honor for the strange power they bring you

Over the past century, typing speed has exceeded people's expectations of physical capabilities, with typists often reaching speeds of 70 words per minute. Although early computer interaction methods were limited by technology, innovations such as local echo improved typing speed and experience, restoring input speeds that matched finger movements. This evolution reflects a deep understanding of human fingers and typing abilities.

Open Source vs the Invisible Hand

The production model of open-source libraries challenges traditional economic theories. Despite characteristics such as non-excludability and zero price, which theoretically should lead to underproduction, there are actually over five million open-source libraries on npm, and the number continues to grow each year. This phenomenon shows that despite the "free rider" problem and the "tragedy of the commons," enough developers still maintain these libraries without pay, forming a complex economic model that combines gift economy, shared infrastructure, and reputation systems.

Snap Unveils Specs, Its $2,200 AR Glasses, and They’re Fugly

Snap has launched its augmented reality glasses, Specs, priced at $2,195, now accepting $200 deposits, with shipping expected this fall. The Specs feature removable lenses, dual lenses that can display content, use proprietary liquid crystal technology, support 16 million colors, and have a battery life of up to 20 hours. The glasses have a unique design aimed at differentiating them from other smart glasses on the market.

Vehicle Motion Cues — a.k.a. Apple’s Weird Anti-Nausea Dots

Apple's 'Vehicle Motion Cues' feature, launched in 2024, uses the device's accelerometer and gyroscope to help users reduce or eliminate motion sickness in moving vehicles. The feature alleviates motion sickness caused by staring at a static display by showing dots on the screen edges that move in sync with the vehicle's motion. Users can configure this feature in the accessibility settings of iOS, iPadOS, and macOS, making it convenient to balance work and life while traveling.

You don’t understand, prices can’t go down

Since 2008, the phenomenon of prices being unable to fall has become increasingly evident, leading to a disconnect between currency and reality. The U.S. economy has grown at an average annual rate of 2.6%, while the stock market has grown at an average annual rate of 12.3%, creating a huge gap that reflects capital continuously shifting from productive activities to bubbles and speculation. If this issue is not taken seriously, it could lead to an economic collapse.

GLM-5.2 is probably the most powerful text-only open weights LLM

Z.ai released GLM-5.2 to its coding plan subscribers on June 13, 2026, and publicly released its weights on June 16. The model has 75.3 billion parameters and a context window of 1 million. GLM-5.2 has become the leading open-weight model in the AI Analysis Index, despite its high output token usage per task, reaching 43,000 tokens. The model ranks second on the Code Arena leaderboard in the Web Development category, behind only Claude Fable 5.

Quoting Charity Majors

In 2025, the economics of code production underwent a disruptive change: code generation became nearly free and instantaneous, turning code from a precious and carefully maintained resource into a disposable commodity that can be regenerated at any time.

Yours Truly on MacBreak Weekly: Is the New Siri AI Good?

In the MacBreak Weekly episode on June 16, 2026, the discussion covered the features of Apple's new Siri and its delayed rollout in the EU, while also mentioning that the iPhone Ultra launch might not be delayed. The show also touched on the limitations private cloud computing imposes on third-party developers, the actual performance of Siri AI, and the UK government's announcement banning social media for users under 16. Additionally, Fox's $25 billion acquisition of Roku's streaming service drew attention.

Yours Truly on The Vergecast: ‘# the **Epic** Story of Markdown’

Markdown was created by John Gruber and has rapidly gained popularity in recent years due to its widespread use in large language model (LLM) systems. Its success lies not only in Markdown itself but in the advantages of plain text files, making human-readable text formatting more efficient and expressive.

<click-to-play> — a still that plays

A progressively enhanced web component called "<click-to-play>" can turn a GIF's link and its first frame into a static placeholder, loading the GIF only when the user clicks it, thus avoiding unnecessary resource consumption. This component aims to improve user experience and is suitable for scenarios where GIFs need to be loaded on demand.

NetNewsWire Status

Brent Simmons has been dedicating his retirement to improving the open-source software NetNewsWire, which was first released in 2002 and became an open-source project in 2018. Users say that after years of use on Mac and iPhone, the software has become an indispensable tool.

Checking In on the iOS Continental Fun-Gap Drift

As European regulations on the iPhone ease, users are gaining more fun and flexibility in app choices and features, such as third-party app stores and the ability to replace default apps. Although these new features are not yet available to users in the United States, many believe that Apple should offer the same options to global users to enhance the overall experience. While the DMA policy has only been in effect for six months, more interesting apps and games may emerge in the future, but for now the focus remains on games like Fortnite.

Summoning the Demon

The current society's pursuit of artificial intelligence reflects humanity's inner evil desires and may lead to societal collapse. Despite doubts about the regulation and oversight of technology, AI is just one tool in an optimization machine that will ultimately bring inevitable consequences. We need to think about the world left after the collapse and build toward that goal.

New in the App Store: Personalized Recommendations

Apple recently announced at its Worldwide Developers Conference that it will introduce personalized recommendation features on the App Store, recommending apps based on users' interests and behaviors. These personalized recommendations will include 'app descriptions' explaining the reason for the recommendation and will be continuously updated based on users' usage and download history. The feature will initially launch in English in the United States and will expand to other regions and languages in the future, while also providing developers with more tools to promote app discovery and user retention.

datasette 1.0a34

Datasette version 1.0a34 adds the ability to insert, edit, and delete data rows within the interface, available in the action items on table pages and row pages. This feature was inspired by Datasette Agent and aims to solve the problem of users being unable to directly edit data in the regular interface.

datasette-tailscale 0.1a0

Released datasette-tailscale 0.1a0, an experimental alpha plugin that allows users to start a local Datasette server via a command and connect to the Tailnet using Tailscale, accessible at http://datasette-preview/. This plugin uses Python bindings for the tailscale-rs library.

Quoting Andrew Singleton

Jenny runs a crematorium, and John's propane company invests $20 billion in her business for a 5% stake. Jenny puts $10 billion into incinerators and uses another $10 billion to buy propane for cremation. John reports that his AI investment generated $10 billion in revenue this quarter and holds a 5% stake in a company valued at $100 billion.

The European Commission Response to Siri AI and the DMA

Apple's decision not to launch 'Siri AI' in the EU is its own choice; the EU's Digital Markets Act (DMA) does not prohibit the introduction of new features. Apple requested an 18-month exemption from interoperability obligations, but this request was denied, leaving EU consumers unable to choose other AI agents. This move may expose approximately 450 million European users to products with limited functionality, further exacerbating Europe's lag in technological innovation.

The WWDC 2026 Keynote and State of the Union on YouTube

Apple's developer app allows users to download local copies of each session, but not the keynote. The reason for this is unclear. The full version of Apple's "State of the Union" is over an hour long, but a 4.5-minute highlight recap has been produced, and it is recommended that users who have not watched the full version at least watch this recap.

Joint Guidance on Vulnerability Naming and Disclosure

The Vulnerability Naming Authority (VNA), in collaboration with the CVE Numbering Authority consortium and the National Telecommunications and Information Administration, has released a unified process for naming, registering, and disclosing vulnerabilities. This process introduces controlled vocabulary, centralized approval registration, and the .vuln top-level domain dedicated to disclosure communications, applicable to publicly disclosed vulnerabilities within the United States. All applications must be submitted through the VNA portal and undergo a five-stage review, with the final decision made by the Naming Committee.

Claude Fable is relentlessly proactive

In a test on June 11, 2026, Claude Fable demonstrated strong initiative, using various techniques to automate problem-solving. By editing Datasette's templates, it successfully triggered keyboard shortcuts and ran a custom web application to collect page information, ultimately helping to identify and fix a UI bug. Although it was downgraded to Opus during the process, Opus was able to continue using Fable's techniques to complete the task and generated a detailed test report.

AI will be massively deflationary

Artificial intelligence will bring about massive deflation. Although some people are critical of Anthropic, they still believe in its marketing, viewing Claude as a self-improving silicon deity. With the proliferation of AI technology, the salaries of knowledge workers will be reassessed, and many industries will be impacted, leading to a decline in wage premiums. The low cost of AI will enable its global adoption, transforming existing social and economic structures.