2026 05 21 Hacker News

2026-05-21 Hacker News Top Stories [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#2026-05-21-hacker-news-top-stories)

1. Google releases Gemini 3.5 Flash, leading in multiple benchmarks with low latency and cost, optimized for large-scale multi-step agent collaboration. Now live in apps and search, speculated to use mixed precision and fewer active parameters.

2. Meta geographically blocks multiple human rights accounts at the request of Saudi and UAE governments, citing cybercrime laws, drawing condemnation from human rights groups and demands for disclosure of legal requests and account restoration.

3. Five major European mobile payment platforms form an alliance to launch a sovereign payment system with interoperability in 2026, initially covering cross-border transfers among 13 countries before expanding to online and offline transactions to localize data and reduce reliance on Visa/Mastercard.

4. Minnesota passes the first comprehensive law banning prediction markets nationwide (including VPN circumvention bans, effective August), with exemptions for insurance, securities, and some weather contracts amid federal regulatory disputes and multiple lawsuits.

5. Google sets Gemini 3.5 Flash as the default AI engine and introduces multimodal, context-retaining conversational search and "Search Agent", but raises concerns among creators about traffic decline and ad model damage.

6. A Tennessee man jailed for 37 days for sharing a Trump meme as a threat, later settled for $835,000 with county and sheriff, highlighting issues of excessive bail, judicial review failures, and free speech protections.

7. GitHub is investigating unauthorized access to ~3,800 internal repositories due to malicious VSCode extensions; no customer data affected but sensitive law enforcement projects involved, exposing supply chain security risks.

8. Qwen announces Qwen Studio, a cross-platform integrated development platform and agent progress, with the community promoting "Omniscience Index" new evaluations to encourage models to say "I don't know" for more comprehensive correctness and hallucination measurement.

9. Railway experiences widespread outage after Google Cloud account suspension, fully restored by May 20, advises some users to manually restart, acknowledges responsibility for vendor choice and may reduce GCP dependency.

10. Texas Tesla lithium refinery accused of discharging 231,000 gallons daily of wastewater containing hexavalent chromium and other pollutants into drainage ditches, with disputes over permits, notifications, regulatory leniency, and sampling methods sparking public health concerns.

1. Gemini 3.5 Flash (Gemini 3.5 Flash) [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#1-gemini-35-flash-gemini-35-flash)

https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5/

This article introduces Google's latest Gemini 3.5 series models, particularly the 3.5 Flash version. Gemini 3.5 aims to combine cutting-edge intelligence with actionable capabilities, significantly enhancing intelligent agents' ability to execute complex tasks. The 3.5 Flash model excels in performance and speed, making it the most powerful agent and coding model to date, capable of efficiently handling long-duration complex tasks with practical application value.

3.5 Flash performs exceptionally well in multiple benchmarks such as Terminal-Bench 2.1, GDPval-AA, and MCP Atlas, and leads in multimodal understanding. Its output speed is four times faster than other leading models, achieving a perfect balance of high intelligence and low latency.

The model is particularly suited for large-scale agent tasks, enabling rapid planning, building, and iteration of solutions, significantly reducing developers' and auditors' work time at lower costs than similar models. Combined with the updated Antigravity platform, 3.5 Flash can deploy collaborative sub-agents to reliably execute multi-step workflows and coding tasks, meeting complex application needs.

Currently, 3.5 Flash is available globally via the Gemini app and Google Search's AI mode. Developers can access it through Google Antigravity, Gemini API, and Android Studio, while enterprise users can connect via the Gemini Enterprise Agent platform. Google also uses the more advanced 3.5 Pro version internally, expected to launch next month. Overall, the Gemini 3.5 series represents a breakthrough in intelligent agent technology, driving wider AI adoption in real-world scenarios.

• * *

https://news.ycombinator.com/item?id=48196570

• It is speculated that the Gemini 3.5 Flash model has around 25-30 billion parameters, with a large total parameter count but fewer active parameters, using mixed precision (FP4/FP8) for efficiency.
• If Gemini 3.5 truly has 200-300 billion parameters and can compete with OpenAI and Anthropic's leading models, it suggests high-performance models may soon run on consumer devices.
• Current small efficient models like DeepSeek V4 Flash, despite high parameter counts, still lag behind the latest GPT Pro and Opus models.
• Opus models have parameters in the trillions, requiring massive scale for performance gains.
• Estimates of model parameters and performance are highly uncertain; actual sizes may be lower.
• Self-hosted open-source models matching Opus 4.7 performance may emerge within a year, given Google's technical lead.
• It's unclear if Opus 4.1/4.5 level breakthroughs can run on consumer devices due to their massive parameter counts.
• While large models can memorize facts, they aren't necessary for meaningful work; smaller models may suffice for most tasks in the future.
• Current hardware like Mac Studio can run hundreds of billions of parameters with near-leading performance.
• Smaller parameter models excel at specific tasks like programming; their capabilities will improve significantly in coming years.
• Leading lab models likely have far fewer parameters than rumored trillions, with optimizations and data constraints affecting size.
• Gemini 3.5 excels in single-inference and coding tasks but performs averagely in long-duration, multi-tool tasks, with a style different from Chinese models.

• * *

https://www.alqst.org/ar/posts/1190

This article reports that Meta has geographically blocked Facebook and Instagram accounts of independent NGOs, researchers, and civil society figures in Saudi Arabia and the UAE, preventing their content from being seen by users in those countries. Since April 30, 2026, Saudi authorities have requested Meta restrict access to ALQST human rights organization, Democratic Diwan, researcher Abdullah Alaoudh, and human rights defender Yahya Assiri's accounts; the UAE has similarly restricted similar accounts. Meta's public reports show over 100 Facebook and Instagram accounts restricted since March 2026, echoing similar blocks on X (formerly Twitter).

These measures are viewed by signatory organizations as arbitrary and discriminatory, severely violating freedom of expression and access to information. Meta claims its actions are based on "local laws" and government requests, primarily citing Saudi and UAE cybercrime laws, which have long been used to suppress dissent and restrict online speech. Restricted content includes coverage of regional geopolitical conflicts and security dynamics. Following the US and Israel's March 2026 attack on Iran, Gulf nations intensified information control.

Signatory organizations question Meta's human rights due diligence before implementing these blocks, demanding transparency on legal requests and human rights assessments, restoration of affected accounts, and detailed explanations of blocked content and applicable laws. They also call for Meta to clarify the role of its Gulf region offices in handling these requests.

In Saudi Arabia and the UAE, governments exercise near-total control over online information, blocking websites like ALQST and Gulf Center for Human Rights focused on human rights and democracy. As a major global social platform, Meta should take greater responsibility to protect free speech and human rights defenders, avoiding becoming a tool for government censorship. However, Meta's actions exacerbate information restrictions, contradicting its public human rights commitments.

Signatories include Access Now, ALQST, US Middle East Rights Commission, and other international and regional human rights groups, jointly urging Meta to adopt transparent, fair measures to safeguard users' free speech and information rights.

• * *

https://news.ycombinator.com/item?id=48206768

• Pursuing short-term growth often leads companies to abandon principles, even treating growth as the sole principle.
• Companies can have principles, but they must be explicitly written into the corporate charter.
• Companies face a choice: comply with government demands or be replaced, which could worsen free speech.
• Individuals can choose not to support companies violating human rights, but as publicly traded entities, companies are legally obligated to pursue profit and growth.
• There are no cases where companies have faced legal liability for sacrificing profits; such claims are often from privileged groups.
• Corporate decisions are executed by people; government requests are communicated to employees for implementation.
• Companies can internally decide on longer-term, ethical strategies rather than short-term gains.
• Comparing corporate behavior to illegal transactions is unreasonable; social responsibility should be prioritized.
• Saudi Arabia and UAE have no extradition treaties but control internet routers, enabling effective censorship.
• UAE's social media restrictions reflect its authoritarian nature despite appeal to expatriates.
• Banning Meta platforms would reduce UAE's attractiveness to expatriates, affecting international exchange.
• Twitter also imposes strict content restrictions, even banning the US president's account.
• Companies prioritize profit over values, a pragmatic but potentially unethical choice.
• Government censorship is morally wrong; corporate compliance is pragmatic but immoral.
• Companies must bear moral responsibility for their actions even under threat.
• Facebook's cooperation with authoritarian regimes to block human rights organizations is unethical.
• Double standards exist when comparing EU and Saudi censorship policies.

3. Goodbye Visa and Mastercard: 130 Million Europeans Switch to Sovereign Payment System [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#3-%e5%91%8a%e5%88%ab-visa-%e5%92%8c-mastercard13-%e4%ba%bf%e6%ac%a7%e6%b4%b2%e4%ba%ba%e8%bd%ac%e5%90%91%e4%b8%bb%e6%9d%83%e6%94%af%e4%bb%98%e7%b3%bb%e7%bb%9f-goodbye-visa-and-mastercard-130m-europeans-switching-to-sovereign-payment)

https://www.lesnumeriques.com/banque-en-ligne/adieu-visa-et-mastercard-130-millions-d-europeens-basculent-vers-un-paiement-100-souverain-des-2026-n250918.html

Five major European mobile payment giants announce an alliance to launch a fully sovereign European payment system in 2026, replacing Visa and Mastercard's dominance. The five platforms include Spain's Bizum, Italy's Bancomat, Portugal's MB WAY, Nordic's Vipps MobilePay, and France's Wero, totaling 130 million active users. By establishing a central interoperability hub, these national payment systems will connect seamlessly, allowing users to transfer money across borders without changing habits.

The system ensures all transaction data bypasses US servers, enhancing Europe's autonomy and data security in digital payments. In 2026, person-to-person transfers will be interoperable across 13 countries, expanding to online and offline payments in 2027. The alliance covers 72% of the EU and Norway's population.

As a pilot, the EuroPA alliance has connected Spain, Portugal, Italy, and Andorra since March 2025, processing €6 million in transactions over the past year, demonstrating strong growth potential. The project reflects Europe's strategic goals in fintech for autonomy and security, addressing concerns over reliance on US payment giants.

• * *

https://news.ycombinator.com/item?id=48207004

• Wero as a EU-wide payment system, similar to the Netherlands' iDeal, allows users to avoid entering card details on merchant sites, redirecting to bank authorization during payment to enhance security and convenience.
• Brazil's PIX payment system is more feature-rich, supporting recurring payments, split payments, financing, and cash withdrawals, with simpler and more intuitive usage.
• The PIX system, controlled by the central bank, reclaims payment control from private institutions, avoiding their restrictions and sanctions.
• Central bank control of payments may pose risks, as governments can monitor and block transactions, potentially infringing on privacy and abusing power.
• Existing private financial institutions, as enforcers of anti-money laundering and sanctions, may block legitimate transactions to avoid legal risks.
• A decentralized payment system allows users to choose different institutions, preventing a single entity from abusing its power to block transactions.
• Marking suspicious transactions differs fundamentally from blocking them; blocking is a preemptive enforcement action with higher risks.
• As payment intermediaries, governments could easily block transactions on a large scale, leading to more severe abuse risks.
• Government recording and monitoring of financial transactions could lead to privacy breaches and power abuse.
• Political pressure may cause private financial institutions to cut off services for certain legal but sensitive businesses, highlighting the need for separation between the state and payment networks.
• All major payment methods (credit cards, iDeal, SWIFT, PayPal, Venmo, etc.) are inherently traceable, as anonymity is not a design goal.
• The independence and political influence of central banks are contentious, potentially affecting the fairness and stability of payment systems.
• Payment systems using central bank infrastructure must still comply with financial regulations and banking secrecy laws, with no fundamental change in information sharing or investigative powers.

• * *

4. Minnesota Becomes First State to Ban Prediction Markets [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#4-%e6%98%8e%e5%b0%bc%e8%8b%8f%e8%be%be%e5%b7%9e%e6%88%90%e4%b8%ba%e9%a6%96%e4%b8%aa%e7%a6%81%e6%ad%a2%e9%a2%84%e6%b5%8b%e5%b8%82%e5%9c%ba%e7%9a%84%e5%b7%9e-minnesota-becomes-first-state-to-ban-prediction-markets)

https://www.npr.org/2026/05/19/nx-s1-5821265/minnesota-ban-prediction-markets

Minnesota has become the first U.S. state to ban prediction market operations. Governor Tim Walz signed legislation making it a felony for prediction market platforms like Kalshi and Polymarket to operate in the state. The law prohibits hosting or promoting prediction markets where users bet on future events, including sports, elections, entertainment, language usage, and international affairs. It also bans using tools like VPNs to circumvent the ban and is set to take effect in August.

The bill was introduced by Democratic Representative Emma Greenman, aiming to have state governments regulate gambling measures to protect public safety and minors. The law includes exceptions for event contracts used for insurance purposes and securities, commodity purchases. The Commodity Futures Trading Commission (CFTC) has filed a lawsuit to block the law, arguing that prediction markets should be exclusively regulated by federal agencies. CFTC Chair Michael Selig stated that Minnesota's law would criminalize legitimate operators and participants overnight, harming farmers and innovators.

Under agricultural industry pressure, the revised law allows trading on weather events. Besides Minnesota, 14 other U.S. states have proposed similar bills restricting prediction markets, with Hawaii and North Carolina seeking a complete ban on the industry. Despite legal uncertainties, prediction market applications continue to grow rapidly, with experts believing the industry has become mainstream and difficult to fully eliminate.

The dispute over state vs. federal regulatory authority has led to over 20 lawsuits. Previously, a Nevada court ruled that Kalshi's sports betting was equivalent to state-regulated gambling, prompting Kalshi to suspend its sports betting operations in the state. The CFTC has filed federal lawsuits against five states, including Arizona, Wisconsin, and New York, seeking to overturn state restrictions on prediction markets.

A Kalshi spokesperson called the ban illegal, comparing it to banning the New York Stock Exchange, and said it would reduce competition, pushing users to offshore platforms. A Polymarket spokesperson stated that Minnesota's ban contradicts the federal government's established regulatory framework for prediction markets. Minnesota allows tribal casinos to operate but prohibits online gambling and sports betting.

Prediction markets allow residents of states banning sports betting to participate in related wagers, as federal law classifies them as "event contracts" rather than gambling. Over 85% of transactions on Kalshi relate to sports events, including high-risk multi-leg bets. Despite large transaction volumes, the industry faces risks of insider trading and manipulation of real-world events. Minnesota Public Radio contributed to this report.

• * *

https://news.ycombinator.com/item?id=48197980

• Minnesota completely bans sports betting, so their rationale for banning prediction markets is more compelling than other states that allow sports betting.
• It's common for laws to distinguish implementation details; allowing sports betting doesn't mean all forms of gambling must be permitted, and regulation and licensing remain necessary.
• Lawmakers sometimes pass new laws explicitly prohibiting already illegal acts to reduce legal ambiguity.
• Prediction markets and sports betting have different risk and regulatory environments; prediction markets may face higher cheating risks, especially for non-sports events.
• It's necessary to distinguish between allowing certain types of gambling and which specific markets are permitted; these issues should be discussed separately.
• Prediction markets could become a key source for factual truth in the future, as financial incentives encourage information accuracy and verification.
• Some prediction markets raise moral and legal issues, such as "murder markets," requiring careful handling and regulation.
• Sports betting also has significant manipulation and insider trading issues; the difference between prediction markets and sports betting in this regard is minimal.
• Discussions should focus on what content is suitable for gambling, rather than equating prediction markets with sports betting.

• * *

5. Google Changes Its Search Box [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#5-%e8%b0%b7%e6%ad%8c%e6%9b%b4%e6%94%b9%e5%85%b6%e6%90%9c%e7%b4%a2%e6%a1%86-google-changes-its-search-box)

https://blog.google/products-and-platforms/products/search/search-io-2026/

This article introduces a series of AI-driven major upgrades in Google's search engine, marking the dawn of a new era for AI search.

First, Google has made its latest Gemini 3.5 Flash model the default engine for AI mode, enhancing search agent and coding performance to meet diverse and complex user queries. Second, it has launched the most significant upgrade to its smart search box in 25 years, supporting multimodal searches via text, images, files, videos, and even Chrome tabs. The search box dynamically expands to intelligently suggest questions based on user intent, helping them express their needs more accurately.

Additionally, the search experience is more conversational; users can ask follow-up questions directly on the AI overview page, with the system maintaining context to provide more relevant links and supporting content, seamlessly across desktop and mobile.

Most notably, the introduction of "search agents." Users can create and manage multiple AI agents that monitor the web 24/7 in the background, including blogs, news, social media, and real-time data (such as finance, shopping, sports), and proactively push comprehensive updates and action recommendations based on specific user needs. For example, when house hunting, the agent continuously scans for matching properties and notifies promptly; when tracking athlete product launches, users receive real-time updates.

Currently, information search agents will be prioritized for Google AI Pro and Ultra subscribers, with future enhancements to further enrich and expand search intelligence capabilities, helping users acquire and utilize information more efficiently.

• * *

https://news.ycombinator.com/item?id=48197370

• Google's AI summaries directly display information, reducing website traffic and making it hard for site owners to benefit.
• Google has tried paying for data, but the payments are far lower than the ad revenue lost from decreased traffic.
• The current monopoly system prevents revenue from reaching content creators, creating a "dam effect."
• Google's ad model faces challenges as AI search reduces user clicks on ads, impacting revenue.
• Traditional internet business models may end, with content creators struggling to earn ad revenue and potentially returning to passion-driven work.
• Google increases ad density to maintain revenue, but AI search may reduce overall ad volume and clicks.
• AI search results may be biased, with partners paying to influence results, and widespread use of user privacy and data.
• Current AI search lacks accuracy and relevance, with misinformation and shallow summaries.
• Some users find AI search provides convenient and accurate summaries with quick access to original sources.
• The convenience of AI search comes with high energy consumption and environmental costs, raising eco-concerns.
• SEO has long been abused, making traditional search harder to find valuable info; AI search is an inevitable trend.

• * *

6. Tennessee Man Jailed 37 Days for Trump Meme Wins Settlement After Lawsuit [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#6-%e7%94%b0%e7%ba%b3%e8%a5%bf%e7%94%b7%e5%ad%90%e5%9b%a0%e7%89%b9%e6%9c%97%e6%99%ae%e8%a1%a8%e6%83%85%e5%8c%85%e8%a2%ab%e5%85%b3%e6%8a%bc-37-%e5%a4%a9%e8%af%89%e8%ae%bc%e8%83%9c%e8%af%89%e8%b5%94%e5%81%bf-tennessee-man-jailed-37-days-for-trump-meme-wins-settlement-after-lawsuit)

https://www.fire.org/news/victory-tennessee-man-jailed-37-days-trump-meme-wins-835000-settlement-after-first-amendment

On May 20, 2026, Larry Bushart, a retired law enforcement official in Tennessee, was wrongfully arrested and jailed for 37 days after sharing a meme containing Trump's remarks. He has now reached an $835,000 settlement with the county and sheriff. In September 2025, after conservative activist Charlie Kirk was shot, Bushart shared a meme quoting Trump's statement "we must overcome it" following a school shooting. The meme referenced a 2024 Iowa shooting, but local police misinterpreted it as a threat to a Perry County high school in Tennessee and withheld key context when applying for an arrest warrant.

Bushart was held on $2 million bail until national attention led to his release. In jail, he lost his post-retirement job, missed his wedding anniversary, and his grandchild's birth. He later partnered with the Foundation for Individual Rights and Expression (FIRE) to file a federal civil rights lawsuit against the police for violating his constitutional rights. A FIRE senior attorney stated that political speech is protected under the First Amendment, and police should not arrest citizens for differing opinions.

FIRE also represents other Tennessee citizens retaliated against for expressing opinions, including civil servants and professors fired for criticizing Kirk. FIRE emphasized that during social unrest, governments must respect free speech or face legal consequences. The organization is dedicated to protecting U.S. citizens' free speech rights and promoting public awareness and respect for these fundamental freedoms.

• * *

https://news.ycombinator.com/item?id=48208502

• The settlement is paid by taxpayers, not individual officers, lacking incentives for police self-regulation and error correction.
• Officers omitted key information when applying for arrest warrants, and judges failed to review effectively, leading to wrongful approvals.
• Judges bear significant responsibility in case handling, while police have limited legal understanding of internet memes.
• The case took too long, delaying justice.
• Bail was excessively high, unreasonably restricting the freedom of the innocent.
• Bail should be based on flight risk or danger, not as punishment.
• Police liability insurance could help reduce abuse of power and improve enforcement standards.
• Deducting settlements from police pensions might incentivize self-discipline but could lead to internal retaliation and toxic culture.
• Current police internal cover-ups and reluctance to report colleagues require careful incentive design.
• Privatized ethics oversight or insurance mechanisms might be more effective than current public systems.
• Insurance costs may ultimately fall on taxpayers, requiring comprehensive solutions.
• Police unions resist reforms, and officers are reluctant to adopt professional licensing and standards like nurses.
• The political stance of leaders toward police unions affects accountability mechanisms.
• Insurance mechanisms could force poorly performing departments to reform due to inability to obtain coverage.

7. GitHub is investigating unauthorized access to their internal repositories [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#7-github-%e6%ad%a3%e5%9c%a8%e8%b0%83%e6%9f%a5%e5%af%b9%e5%85%b6%e5%86%85%e9%83%a8%e4%bb%a3%e7%a0%81%e5%ba%93%e7%9a%84%e6%9c%aa%e6%8e%88%e6%9d%83%e8%ae%bf%e9%97%ae-github-is-investigating-unauthorized-access-to-their-internal-repositories)

https://twitter.com/github/status/2056884788179726685

This webpage prompts users that their browser has JavaScript disabled, suggesting enabling JavaScript or switching to a supported browser to continue using the X platform. The page provides links to the help center, terms of service, privacy policy, and cookie policy.

The webpage also displays an announcement from GitHub's official account stating they are investigating unauthorized access to their internal repositories. There is currently no evidence that customer information (such as enterprises, organizations, and code repositories) has been affected, but GitHub is closely monitoring infrastructure to prevent further activity.

Additionally, the page includes X platform login and registration options, encouraging new users to sign up for personalized timelines. It also displays trending topics in the U.S. region, such as musician Dom Smith, band The Clash, and political figure Kilmeade.

Overall, this webpage serves as a social media platform login and information announcement page, emphasizing security incident notifications and user login/registration features while showcasing real-time trending topics.

• * *

https://news.ycombinator.com/item?id=48201316

• GitHub confirmed approximately 3,800 internal repositories were leaked via a malicious VSCode extension, with severe impact.
• Currently, only internal repositories have been confirmed compromised; no evidence indicates customer external information was affected.
• Leaked repositories include law enforcement-related files and departments, indicating GitHub has a legitimate process for handling law enforcement requests.
• The leaked file list was published by the hacker group TeamPCP, containing multiple sensitive projects.
• Some repository names reference specific groups or organizations, reflecting GitHub's diverse internal teams and projects.
• GitHub's statement is formal, avoiding direct acknowledgment of a breach, showing caution.
• The incident may trigger regulatory requirements for timely reporting, requiring GitHub to notify authorities within a specified timeframe.
• Investors have shown limited reaction to the incident, as large-scale security breaches typically have minimal market impact.
• Some comments joked about the use of the term "directional," expressing differing views on the incident's severity.
• Some suggested using static analysis and firewalls to enhance supply chain security.

• * *

8. Qwen3.7-Max: The Agent Frontier [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#8-qwen37-max%e6%99%ba%e8%83%bd%e4%bb%a3%e7%90%86%e5%89%8d%e6%b2%bf-qwen37-max-the-agent-frontier)

https://qwen.ai/blog?id=qwen3.7

This webpage primarily introduces Qwen Studio and its related products and services. Qwen Studio is a platform integrating various features, including code editing, research tools, and API interfaces, designed to provide users with a convenient development and research experience. The page mentions a current content loading issue, advising users to refresh the page or contact technical support via email for assistance.

The webpage also displays the supported operating systems and devices for Qwen Studio, including Web, iOS, Android, macOS, and Windows, highlighting its cross-platform capabilities. The page footer includes relevant legal terms, privacy policy, usage policy, cookie notice, and training data summary to ensure users understand the platform's compliance and data usage.

Overall, this webpage serves as the official introduction and service entry for Qwen Studio, emphasizing its multifunctionality, multi-platform support, and user service guarantees.

• * *

https://news.ycombinator.com/item?id=48205626

• The AA-omniscience model's non-hallucination rate is at a new high, surpassing Opus 4.7, Gemini 3.1 Pro, and GPT5.5.
• Counting only incorrect answers for hallucination rates may lead models to avoid hallucinations by refusing to answer.
• The term "hallucination" was introduced because large models often fabricate answers; refusing to answer can enhance model credibility.
• The Omniscience index comprehensively evaluates model performance by considering correct answers, hallucinations, and refusals.
• Models that can answer "I don't know" are more trustworthy than those that always provide answers.
• Early models had the ability to assess answer probabilities and should have refused to answer or explicitly warned of potential hallucinations at low confidence levels.
• Some Chinese models perform poorly in knowledge accuracy and hallucination rates.
• Models vary significantly in token efficiency; some require more tokens to achieve the same answer quality.
• Test data may contain human errors; a perfect non-hallucination rate doesn't guarantee absolute model correctness.
• The definition and judgment of hallucinations depend on the tester's knowledge framework, which may have cognitive limitations.
• It remains unknown whether current models can surpass human cognitive limits to understand the world more accurately.
• Evaluation questions are relatively clear, making it easier to judge the correctness of model responses.
• Run speed and hardware configuration significantly impact model usability; mixture-of-experts models have advantages in speed and memory usage.
• 27B models perform better on code tasks but have higher hardware requirements.
• Quantization techniques affect model performance and speed; dynamic quantization allows for trade-offs between precision and efficiency.

• * *

9. Incident Report: Google Cloud Caused Railway Service Outage [Resolved] [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#9-%e4%ba%8b%e4%bb%b6%e6%8a%a5%e5%91%8agoogle-cloud-%e5%af%bc%e8%87%b4-railway-%e6%9c%8d%e5%8a%a1%e4%b8%ad%e6%96%ad%e5%b7%b2%e8%a7%a3%e5%86%b3-incident-report-railway-blocked-by-google-cloud-resolved)

https://status.railway.com/incident/I23M92U0

This webpage is Railway's incident update and recovery report regarding a service outage caused by a Google Cloud account issue on May 19, 2026. The incident was investigated starting at 22:29 on May 19, with users experiencing login failures, inability to access dashboards, and various error messages. Subsequently, the Railway team confirmed that Google Cloud had blocked their account, causing partial service unavailability, and escalated the issue directly with Google.

Over the following hours, the team gradually restored access to Google Cloud infrastructure, with some services and workloads coming back online, though intermittent network and deployment issues persisted. To prevent infrastructure overload, builds for non-enterprise users were temporarily restricted. By 07:57 on May 20, Railway services were fully restored, with the system automatically redeploying workloads detected as anomalous; users can manually trigger redeployment if issues persist.

The detailed cause of the incident has been confirmed, with a more comprehensive incident analysis report to follow. Railway apologized for the outage and provided relevant FAQs and community support links for users to stay updated and seek assistance.

• * *

https://news.ycombinator.com/item?id=48201484

• The Railway incident is resolved, but some users still face service unavailability and need to manually restart to recover.
• Railway acknowledges responsibility for vendor selection; users care about product availability, so the responsibility lies with Railway.
• Some argue that suing Google for damages is possible, but Google's terms are strict, making litigation success unlikely.
• Questions arose whether Google automatically suspended the account due to payment issues; Railway stated a dedicated account manager was handling it.
• Google may offer apologies and account credits; Railway might consider reducing reliance on GCP.
• Google's terms of service are typically broad, allowing account suspension without warning, limiting user rights.
• A multi-cloud strategy doesn't require deploying multiple clouds simultaneously; demonstrating the leverage of multi-cloud options suffices.
• AWS and Azure also experience outages, but GCP's automatic account suspension incident drew particular attention.
• AWS us-east-1 region outages have widespread impact, as many services depend on it, causing cascading effects.
• Some users note that AWS control panels occasionally fail, but data panels remain stable, limiting impact.
• Large enterprises heavily rely on the us-east-1 region, causing broader impact during regional outages, which are hard to fully avoid.

• * *

10. Tesla's Lithium Refinery Discharges 231,000 Gallons of Polluted Wastewater Daily [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#10-%e7%89%b9%e6%96%af%e6%8b%89%e9%94%82%e7%b2%be%e7%82%bc%e5%8e%82%e6%af%8f%e5%a4%a9%e6%8e%92%e6%94%be-231-%e4%b8%87%e5%8a%a0%e4%bb%91%e6%b1%a1%e6%9f%93%e5%ba%9f%e6%b0%b4-teslas-lithium-refinery-discharges-231000-gallons-of-polluted-wastewater-a-day)

https://www.autonocion.com/us/tesla-lithium-refinery-texas/

In January 2026, Nueces County drainage district staff discovered an unfamiliar pipe discharging black wastewater into a canal they manage during routine inspections. Confirmed to belong to Tesla, the pipe discharges wastewater from its $1 billion lithium refinery, which began operations in December 2024. The facility claims to use an "acid-free cleaning process" with sand and limestone as primary byproducts, but the drainage district was not informed of the daily discharge of 231,000 gallons of treated wastewater through their facilities.

The Texas Commission on Environmental Quality (TCEQ) quietly approved Tesla's wastewater discharge permit in January 2025, allowing discharge into an unnamed canal that flows into Baffin Bay. However, the permit did not explicitly authorize Tesla to use public or private land for the canal, and the drainage district was not notified of the permit's existence. Drainage district staff discovered the pipe during field inspections and filed a complaint with TCEQ.

TCEQ conducted routine pollutant tests on the discharged water in February 2026, which met permit standards, but did not test for heavy metals or lithium. The drainage district later hired an independent lab for 24-hour sampling, detecting harmful substances including hexavalent chromium, arsenic, strontium, lithium, and vanadium. Hexavalent chromium is a known human carcinogen, and arsenic levels, though below drinking water standards, were present. Other elements like manganese, iron, and phosphorus also showed industrial discharge characteristics, posing potential threats to ecosystems and human health.

Drainage district engineers recommended Tesla build multi-stage reverse osmosis facilities to remove heavy metals and warned local residents to avoid the canal. Increased salinity has killed canal bank grasses, raising risks of soil erosion and reduced drainage capacity. The drainage district's attorney issued a cease-and-desist notice to Tesla, demanding a meeting to discuss the issue.

This incident highlights contradictions between environmental commitments and regulatory gaps in the U.S. EV supply chain, exposing issues with Tesla's lithium refinery wastewater discharge oversight and transparency, which have not been widely covered by mainstream automotive media.

• * *

https://news.ycombinator.com/item?id=48198551

• Tesla obtained a discharge permit allowing treated wastewater into a designated unnamed canal, but disputes exist over whether they have the right to use public or private property for the canal.
• The canal management authority was not informed of the permit; responsibility lies with the issuing agency, TCEQ.
• Tesla questioned the testing methodology, arguing samples should be taken at the discharge point rather than downstream to avoid interference from other pollution sources.
• While permitting and approval processes are seen as bureaucratic, they are necessary checks to ensure emissions meet industry standards and safety.
• The discharge permit allows wastewater into the canal, but using the canal itself may require additional permits, involving overlapping jurisdictions and management authorities.
• The "unnamed canal" mentioned in the permit refers to a specific canal, not a vague one; the dispute centers on ownership and usage rights of the canal.
• Texas regulations are relatively lenient, and regulatory agencies may overlook environmental and public health concerns for economic interests.
• Discharge permits are issued by water quality regulators, but canal management falls under the drainage district, leading to coordination difficulties due to overlapping authorities.
• Tesla and the drainage district disagree on water testing methods, affecting pollution level assessments.

Hacker News Featured Comments and Translation [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#hacker-news-%e7%b2%be%e5%bd%a9%e8%af%84%e8%ae%ba%e5%8f%8a%e7%bf%bb%e8%af%91)

GitHub confirms breach of 3,800 repos via malicious… [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#github-confirms-breach-of-3800-repos-via-maliciou)

https://news.ycombinator.com/item?id=48214112

If only the companies behind VSCode, NPM, and GitHub could collaborate to find a solution.

psadauskas

If only the companies behind VSCode, NPM, and GitHub could come together to figure out a solution.

• * *

Goodbye Visa and Mastercard: 130M Europeans switch… [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#goodbye-visa-and-mastercard-130m-europeans-switch)

https://news.ycombinator.com/item?id=48207532

Wero is essentially an EU-wide version of the Dutch iDeal system, which I consider the gold standard for how internet payments should work. I shouldn't have to enter card details on merchant sites (which is unsafe). Instead, payments should redirect me to my bank, where I authorize them through my bank's security system. I've always been annoyed by having to input sensitive card information on various merchant sites. I hope Wero gains broader support now that it's being used across the EU.

mcv

Wero is essentially an EU-wide version of the Dutch iDeal system, which I consider the gold standard for how internet payments should work. I shouldn't have to enter card details on merchant sites (which is unsafe). Instead, payments should redirect me to my bank, where I authorize them through my bank's security system. I've always been annoyed by having to input sensitive card information on various merchant sites. I hope Wero gains broader support now that it's being used across the EU.

• * *

Google changes its search box [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#google-changes-its-search-box)

https://news.ycombinator.com/item?id=48197769

This raises a fascinating question for anyone running a website: if Google is just crawling my sites and presenting AI summaries on their platform, what exactly do I gain by allowing Googlebot to crawl my site?

ekidd

This raises a fascinating question for anyone running a website: if Google is just crawling my sites and presenting AI summaries on their platform, what exactly do I gain by allowing Googlebot to crawl my site?

• * *

Map of Metal [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#map-of-metal)

https://news.ycombinator.com/item?id=48206213

Didn't expect to see something I made on HN while my wife was trying to find something to watch on TV.

Let me talk about the site in case anyone's interested. I made it with a friend studying multimedia. He handled the data, and I did the coding. Took about a week or two.

The site was originally built with Flash (remember that?). A few years ago, I ported it to HTML5. It still has that Flash vibe, I think. I posted the code to GitHub when I ported it, mostly to keep it alive for old times' sake.

As for mobile support, I planned to do it but got sidetracked building a custom WebGL map renderer because phone performance was poor. However, I never finished—life always finds a way to get in the way... I have some mobile designs lying around.

Another issue was when I first built the site: YouTube didn't really have many ads back then, just small text ads, and you could embed the player really tiny. So it worked better. In the original Flash version, I actually hid the video player. But that got the site blacklisted by YouTube. I asked a Google engineer on a dev forum to intercede, and they removed the block. Those were different times—back when Google was a different beast, and you could chat with real people online, and developer communities were much smaller.

I have a sketch of a much larger map in my notebook. It includes many more subgenres and interconnected elements like historical events, etc. It's huge when unfolded—about 2x1.5 meters or something ridiculous.

I miss those days when the web was full of weird and experimental stuff. I grew up with Newgrounds and Geocities, and I'm sure all that stuff is still out there, buried under a mountain of SEO-optimized garbage.

pjgalbraith

• * *

An OpenAI model has disproved a central conjecture… [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#an-openai-model-has-disproved-a-central-conjecture)

https://news.ycombinator.com/item?id=48213071

To the "LLMs just interpolate their training data" crowd:

Ayer, and in a different way early Wittgenstein, held that mathematical truths don’t report new facts about the world. Proofs unfold what is already implicit in axioms, definitions, symbols, and rules.

I think that idea is deeply fascinating, AND have no problem that we still credit mathematicians with discoveries.

So either “recombining existing material” isn’t disqualifying, or a lot of Fields Medals need to be returned.

m-hodges

To the "LLMs just interpolate their training data" crowd:

Ayer, and in a different way early Wittgenstein, held that mathematical truths don’t report new facts about the world. Proofs unfold what is already implicit in axioms, definitions, symbols, and rules.

I think that idea is deeply fascinating, AND have no problem that we still credit mathematicians with discoveries.

So either “recombining existing material” isn’t disqualifying, or a lot of Fields Medals need to be returned.

• * *

Tennessee man jailed 37 days for Trump meme wins s… [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#tennessee-man-jailed-37-days-for-trump-meme-wins-s)

https://news.ycombinator.com/item?id=48209448

The sheriff that arrested him should face criminal charges for misuse of authority. That he doesn’t reflects a structural weakness in US law. In most European legal systems a law enforcement officer overstepping his legal authority would face criminal charges for it.

contubernio

The sheriff who arrested him should face criminal charges for abuse of authority. His lack of prosecution reflects a structural weakness in US law. In most European legal systems, a law enforcement officer exceeding legal authority would face criminal charges.

• * *

Goodbye Visa and Mastercard: 130M Europeans switch… [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#goodbye-visa-and-mastercard-130m-europeans-switch-1)

https://news.ycombinator.com/item?id=48207417

A Frenchman using Wero will be able to transfer money to a Spanish friend on Bizum, with the same simplicity as a domestic payment.

Have you seen the new money app? It’s on Tubu. It’s on Weeno. I’m on Dippy but my friend is on Poob. Poob has it for you.

petcat

A Frenchman using Wero will be able to transfer money to a Spanish friend on Bizum with the same simplicity as a domestic payment.

Have you seen the new payment app? It’s on Tubu. It’s on Weeno. I’m on Dippy but my friend is on Poob. Poob has it for you.

• * *

Gemini 3.5 Flash [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#gemini-35-flash)

https://news.ycombinator.com/item?id=48202262

For those who would like to know the total and active parameter count of this model: even though Google doesn’t disclose the model technicals, we can infer them within relatively tight margins based on what we do know.

We know they serve the model on TPU 8i, which we have plenty of hard specs for (so we know the key constraints: total memory and bandwidth and compute flops). We can also set a ceiling on the compute complexity and memory demand of the model based on knowing they will be at least as efficient as what is disclosed in the Deepseek V4 Technical Report.

We can also assume that the model was explicitly built to run efficiently in a RadixAttention style batched serving scenario on a single TPU 8i (so no tensor parallelism, etc. to avoid unnecessary overheads… Google explicitly designed the 8th-generation inference architecture to eliminate the need for tensor sharding on mid-sized models).

We know Google intends to serve this model at a floor speed of around 280 tok/s too.

Putting all these pieces together, we can confidently say this model is ~250-300B total, and 10-16B active parameters. Likely mostly FP4 with FP8 where it matters most.

Visual:

┌────────────────────────────────────────────────────────┐ │ TPU 8i VRAM (288 GB) │ ├───────────────────────────┬────────────────────────────┤ │ Static Model Weights │ Dynamic Allocations & │ │ (250B - 300B @ Mixed │ Compressed KV Caches │ │ FP4/FP8) │ (RadixAttention / SRAM) │ │ ~110 GB - 150 GB │ ~138 GB - 178 GB │ └───────────────────────────┴────────────────────────────┘ I do model serving optimization work. This is napkin math.

Edit: There’s one factor I under-rated in my initial estimate… TurboQuant. This is a compute to KV memory use tradeoff. It’s plausible with TurboQuant at a quality-neutral setting they’ve gotten the model up to 400B with similar economics. This is a variable effecting concurrency and the the way they decided total model size was likely based on what they see for the average user’s average KV cache depth in real-world usage.

easygenes

For those who want to know the total and active parameter count of this model: although Google doesn’t disclose the technical details, we can infer them within relatively tight margins based on known information.

We know the model is served on TPU 8i, for which we have extensive hard specs (so we know the key constraints: total memory, bandwidth, and compute flops). We can also set an upper bound on the model’s compute complexity and memory demand based on its efficiency being at least as good as disclosed in the Deepseek V4 Technical Report.

We can also assume the model was explicitly designed to run efficiently in a RadixAttention-style batched serving scenario on a single TPU 8i (so no tensor parallelism, etc., to avoid unnecessary overheads… Google explicitly designed the 8th-generation inference architecture to eliminate tensor sharding for mid-sized models).

We also know Google intends to serve this model at a minimum speed of around 280 tokens per second.

Putting all these pieces together, we can confidently say the model has ~250-300B total parameters and 10-16B active parameters. Likely mostly FP4 with FP8 where it matters most.

Visual:

┌────────────────────────────────────────────────────────┐ │ TPU 8i VRAM (288 GB) │ ├───────────────────────────┬────────────────────────────┤ │ Static Model Weights │ Dynamic Allocations & │ │ (250B - 300B @ Mixed │ Compressed KV Caches │ │ FP4/FP8) │ (RadixAttention / SRAM) │ │ ~110 GB - 150 GB │ ~138 GB - 178 GB │ └───────────────────────────┴────────────────────────────┘

I work on model serving optimization. This is napkin math.

Edit: One factor I underestimated in my initial estimate… TurboQuant. This is a compute-to-KV-memory tradeoff. With TurboQuant at a quality-neutral setting, they may have scaled the model to 400B while maintaining similar economics. This variable affects concurrency, and the total model size was likely determined based on observed average KV cache depth for real-world users.

• * *

Incident Report: May 19, 2026 – GCP Account Suspen… [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#incident-report-may-19-2026--gcp-account-suspen)

https://news.ycombinator.com/item?id=48211323

“Finally, we are in planning to remove Google Cloud services from our data plane’s hot path, and keeping them only for secondary/failover.”

That’s pretty clear. Google can no longer be trusted as a B2B service provider.

Animats

“Finally, we are planning to remove Google Cloud services from our data plane’s hot path, keeping them only for secondary/failover.”

That’s clear. Google can no longer be trusted as a B2B service provider.

• * *

Gemini CLI will stop working from June 18, 2026 [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#gemini-cli-will-stop-working-from-june-18-2026)

https://news.ycombinator.com/item?id=48201329

Google really can’t help themselves but to have some internal re-org kill off a public thing people are actively using. It’s honestly impressive how consistent they are.

silverlight

Google really can’t help but to have internal reorganizations kill off public features people are actively using. It’s honestly impressive how consistent they are.

• * *

https://news.ycombinator.com/item?id=48203416

Do they know what the attackers were after? Maybe they were just trying to help fix the availability problems.

jms703

Do they know what the attackers were after? Maybe they were just trying to help fix the availability issues.

• * *

Google changes its search box [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#google-changes-its-search-box-1)

https://news.ycombinator.com/item?id=48199476

What scares me about this new AI mode thingy

What scares me is the rampant inaccuracy. In my experience, the AI responses are wrong about 65% of the time. I just did a search today about an error talking about a disconnected link between apps, and Google AI result summary told me that the error was related to my pulling a USB drive too quickly in windows. The ONLY word similar to my query and that AI response was the word “disconnect”. Everything else was clearly about the SaaS apps.

I have people coming to me, asking me questions, then telling my Google told them something else, so now I have to waste time convincing them that it’s wrong. Over the past 2 years AI has done nothing for me but complicate my work life.

And of course, this could be because the model is crap, but it could be because they want me to keep refining my query over and over for more ad views. Either way, it’s a terrible experience.

burnte

Incident Report: Railway Blocked by Google Cloud [… [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#incident-report-railway-blocked-by-google-cloud-)

https://news.ycombinator.com/item?id=48201884

It has been 0 days since GCP has taken down a startup (again).

You see this at least once a year. Never heard of this from AWS or Azure.

In all seriousness, this is why we don’t use them. They have the most ergonomic cloud of the big three, then absolutely murder it by having this kind of reputation.

dangoodmanUT

GCP has taken down a startup again, and it's been 0 days.

You see this at least once a year. Never heard of this from AWS or Azure.

In all seriousness, this is why we don't use them. They have the most ergonomic cloud among the big three, but they completely ruin it with this kind of reputation.

• * *

Tennessee man jailed 37 days for Trump meme wins s… [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#tennessee-man-jailed-37-days-for-trump-meme-wins-s-1)

https://news.ycombinator.com/item?id=48210587

The fact that taxpayers and not the police themselves have to pay the settlement is the worst part of this.

Every settlement against the police should be taken from their pension fund. This is something I’ve been advocating for decades now, because it creates an incentive not to do things like this. Right now, good cops don’t patrol bad cops because it won’t affect them. By aligning the incentives right, it will mean good cops will force out the bad cops quickly.

freediddy

The worst part is that taxpayers, not the police themselves, have to pay the settlement.

I've been advocating for decades that all police settlements should come from their pension funds. This creates an incentive to avoid such actions. Currently, good cops don't monitor bad cops because it doesn't affect them. Properly aligning incentives would make good cops quickly eliminate the bad ones.

• * *

Meta blocks human rights accounts from reaching au… [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#meta-blocks-human-rights-accounts-from-reaching-au)

https://news.ycombinator.com/item?id=48207311

Social media companies post record earnings year after year from their ads business while increasingly proving to be harmful to society. They do the bare minimum in terms of content moderation and bots while priming the algorithms to maximize revenue. The good ol’ privatized profits, socialized harm model.

In a just world, would social media platforms be taxed higher on corporate revenue and how would that pan out? Maybe we’ll be left with small federated platforms without algorithms and ads.

0x5FC3

Social media companies report record earnings year after year from ads while becoming increasingly harmful to society. They barely do content moderation and bot control, instead optimizing algorithms for maximum revenue. Classic privatized profits, socialized harm.

In a fair world, would social media platforms face higher corporate taxes? What would that look like? Perhaps we'd end up with small federated platforms without algorithms or ads.

• * *

Google changes its search box [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#google-changes-its-search-box-2)

https://news.ycombinator.com/item?id=48197660

I don’t trust facts from LLMs. When I am searching for something, I usually want to find primary sources. As soon as a number is involved, I do my best to not even look at the AI output.

Even though the result is often good and combines information from multiple sources, it can also get things wrong by combining information from different eras or just plain outdated advice. AFAICT, without primary sources, the result is for entertainment purposes only.

imoverclocked

I don't trust facts from LLMs. When searching for something, I usually want primary sources. If numbers are involved, I avoid looking at AI output entirely.

Even though results often combine info from multiple sources, they can mix up data from different eras or give outdated advice. Without primary sources, the results are just for entertainment.

• * *

Apple unveils new accessibility features [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#apple-unveils-new-accessibility-features)

https://news.ycombinator.com/item?id=48195239

Apple loves to stealth test new tech in full public view by sneaking it into relatively mundane places, so debuting agentic AI via accessibility is very on brand.

A few other examples:

• The Touch Bar was much more than an OLED strip, it was Apple’s first move in the transition to Apple Silicon on macs. The Apple T1 chip in the 2016 Touch Bar MacBooks was the first solely Apple-designed processor to appear in a Mac and took over several responsibilities away from intel chipsets like power management, fans, sleep/wake, access to the camera & mic, and the secure enclave powering touch ID. Then the T2 added encryption of the SSD, audio management, image processing for the camera, and prevented tampering with the boot process

• The iPhone 3G shipped with a Liquidmetal SIM eject tool, which is made from a strong custom metal alloy which is “practically unbendable by hand unless you want to hurt or cut your fingers.” Although Apple hasn’t released anything with the alloy since then, now nearly 20 years later Apple is rumored to be using liquid metal in their upcoming foldable iPhone.

• RealityKit has supported 3D scanning and many other cool AR capabilities for years, which only made sense after the Apple Vision Pro was released.

JeremyHerrman

Apple has a habit of quietly integrating new technologies into seemingly mundane features to test them in the public eye, so introducing autonomous AI for the first time through accessibility features is very much in line with Apple's style.

Here are a few more examples:

• The Touch Bar isn't just an OLED strip; it was Apple's first step toward transitioning Macs to Apple Silicon. The Apple T1 chip in the 2016 Touch Bar MacBook was the first Mac processor fully designed by Apple, handling tasks previously managed by Intel chipsets, such as power management, fan control, sleep/wake functions, camera and microphone access, and the Secure Enclave for Touch ID. The subsequent T2 chip added SSD encryption, audio management, camera image processing, and prevented boot process tampering.

• The iPhone 3G came with a liquid metal SIM ejector tool made from a high-strength custom metal alloy that's "nearly impossible to bend by hand unless you want to hurt your fingers or cut yourself." Although Apple hasn't released any products with this alloy since then, nearly 20 years later, rumors suggest it may be used in the upcoming foldable iPhone.

• RealityKit has supported 3D scanning and many other cool AR capabilities for years, which only made sense after the Apple Vision Pro was released.

• * *

Gemini CLI will stop working from June 18, 2026 [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#gemini-cli-will-stop-working-from-june-18-2026-1)

https://news.ycombinator.com/item?id=48203833

How does anyone internally at Google justify these decisions?

Even if there are competing implementations, in terms of brand recognition, I feel like “Gemini” is more closely associated with Google than “Antigravity”. Why pick the more obscure option?!

Perhaps they felt the sentiment on Gemini CLI was beyond repair, but surely there must be some voice on the inside saying “developers will never adopt our products if we keep killing them”.

crakhamster01

Who at Google internally justifies these decisions?

Even with competing implementations, in terms of brand recognition, "Gemini" is more closely associated with Google than "Antigravity". Why choose the more obscure option?

Perhaps they felt the sentiment around Gemini CLI was beyond repair, but surely someone inside must be saying, "Developers will never adopt our products if we keep killing them off."

• * *

Google changes its search box [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#google-changes-its-search-box-3)

https://news.ycombinator.com/item?id=48197951

A couple of years back I worked with a company which maintained specific data which was the main traffic driver on that page. Google approached them and wanted to pay for the rights to get the data and display it on top of the search results, a feature which was fairly new back then.

This was an interesting dilemma because it was very clear that the money was way less than the loss in ad revenue due to traffic drop, but it was also clear that if we wouldn’t take the deal, a more desperate competitor would, which would result in the same traffic loss but without the extra google money. So the company took the deal.

History repeats itself here, with the difference that instead of paying for the data, the ai crawlers simply take it for free.

pflenker

A few years ago, I worked with a company that maintained specific data which was the main traffic driver for that page. Google approached them to pay for the rights to use the data and display it at the top of search results, a feature that was relatively new at the time.

This presented an interesting dilemma because it was clear that the money offered was far less than the ad revenue loss from the traffic drop, but also clear that if we didn't take the deal, a more desperate competitor would, resulting in the same traffic loss without the additional Google money. So the company accepted the deal.

History is repeating itself here, but this time, AI crawlers simply take the data for free instead of paying for it.

• * *

Gemini 3.5 Flash [#](https://supertechfans.com/cn/post/2026-05-21-HackerNews/#gemini-35-flash-1)

https://news.ycombinator.com/item?id=48199413

This understates the cost increase. 3.5 Flash also uses more tokens. artificialanalysis.ai shows these difference to run the whole eval, which I think is more realistic pricing:

Gemini 2.5 flash (27 score): $172 (1.0x)

Gemini 2.5 pro (35 score): $649 (3.8x)

Gemini 3.0 Flash (46 score): $278 (1.6x)

Gemini 3.5 Flash (55 score): $1,552 (9.0x or 2.4x compared to 2.5 pro)

This is a massive price increase… 5.6x compared to Gemini 3.0 Flash

jl

This underestimates the cost increase. 3.5 Flash also uses more tokens. artificialanalysis.ai shows these differences for the entire evaluation, which I believe reflects more realistic pricing:

Gemini 2.5 Flash (27 score): $172 (1.0x)

Gemini 2.5 Pro (35 score): $649 (3.8x)

Gemini 3.0 Flash (46 score): $278 (1.6x)

Gemini 3.5 Flash (55 score): $1,552 (9.0x or 2.4x compared to 2.5 Pro)

This is a massive price increase... 5.6x compared to Gemini 3.0 Flash