2026 06 12 HackerNews

2026 06 12 HackerNews | SuperTechFans

2026-06-12 Hacker News Top Stories #

πFS is an April Fools' joke file system that claims data is not stored on the hard drive but instead stored in the decimal places of π, but in reality, it is extremely slow and serves only as a proof of concept. Homebrew 6.0.0 has been released, introducing tap trust mechanisms, Linux sandbox support, parallel installation, and other improvements, and fixing security vulnerabilities. Niantic has used scan data from Pokémon Go players to train a visual localization system and plans to apply it to military drone navigation in collaboration with a U.S. defense contractor, raising ethical concerns. An AI agent caused chaos in Fedora and upstream projects, and the community suspects it was a prelude to a supply chain attack; the account has since been revoked. A farmer in Texas donated land worth $10 to build a park, but the city government sold it for $10 million to a data center developer, prompting residents to sue. The Windows version of Claude Desktop automatically creates a 1.8 GB Hyper-V virtual machine every time it starts, even if it's only used for chatting, causing system lag. MiMo Code is an open-source AI programming assistant with cross-session persistent memory, supporting multiple agents and any LLM, and is released under the MIT license. In May 2026, solar power generation in the United States exceeded coal for the first time, becoming the third largest source of electricity, mainly driven by economic factors, despite policy support for coal. The author criticizes AI vendors for using vanity metrics such as "lines of code" to package productivity, but actual research shows that AI assistance has not significantly improved productivity, and calls for the use of reliable metrics. The article humorously reviews the naming evolution of Anthropic models, listing a series of fictional literary names, mocking naming inflation and industry jargon.

1. πFS (πFS) #

https://github.com/philipl/pifs

πfs is a "free data" file system that claims not to store data on the hard drive but instead stores it in π. Its core premise is that π is considered a normal number, so all possible finite digit sequences will appear in the decimal part of π, meaning that any file can be found in π. πfs searches for each byte individually in π and stores the location of the file in π as metadata in a separate metadata directory.

Building requires autoconf, automake, and libfuse, using the standard autotools process (./autogen.sh → ./configure → make → make install). The mount command is: πfs -o mdd=<metadata directory> <mount point>.

This project is an April Fools' joke, with extremely slow performance (storing a 400-line text file takes five minutes), but the author mentions possible optimization directions (variable-length search, arithmetic coding, parallel search, cloud π search, etc.). The project README also directs users to follow updates on inferencefs.

HN热度 933 points | 评论 201 comments | 作者：helterskelter | 1 day ago #

https://news.ycombinator.com/item?id=48480978

• Using the Library of Babel as a data compression tool requires the address itself to have the same amount of information as the data, so it cannot truly compress, but it is an interesting thought experiment.

• LLM is a form of lossy compression, achieving what similar tools could not, but it requires a lot of groundwork and is related to the concept of language compression.

• The video "Reinventing Entropy Compression is Intelligence" by 3Blue1Brown explores the connection between compression and intelligence.

• Most usernames mentioned in the comments correspond to various constructed languages (conlangs), forming a kind of language club.

• Science is an extreme form of compression, for example, Newtonian mechanics can explain a large number of phenomena with a small amount of text, but it is a highly lossy compression.

• Modern LLMs (such as GPT-2) combined with arithmetic coding can achieve lossless text compression, reaching a compression rate of about 1 bit per byte; more advanced models perform even better.

• LLM is a strange and interesting result of applying traditional lossy compression concepts (such as audio, image, and video) to the text domain.

• When the temperature of the LLM is set to 0.0, it is almost like a key-value store, but finding the correct key requires some effort.

• This "data and address are the same size" dilemma reminds people of Borges' parable of the "Map of the Empire."

• Unconditional security (such as one-time pad) requires a key length at least equal to the message length, while other encryption is based on unproven assumptions.

• Scientific laws as compression are highly lossy, for example, "the old lady fell" may not only be gravity, but also involve more details.

• The implicit forces AI learns when simulating planetary orbits are absurd, different from the concise compression of Newton's laws.

2. Show HN: Homebrew 6.0.0 #

https://brew.sh/2026/06/11/homebrew-6.0.0/

Homebrew 6.0.0 has been officially released. Major updates include: introducing the tap trust mechanism, requiring third-party taps to be explicitly trusted before running code; enabling the internal JSON API by default, speeding up updates; Linux version adds sandbox support (based on Bubblewrap); improving default settings based on user surveys, such as displaying dependency summaries and confirming in developer mode by default when running brew install; brew bundle supports parallel installation, npm/krew extensions, Windows winget, etc.; performance optimization, brew leaves is about 30% faster. Preliminary support for macOS 27 (Golden Gate). At the same time, three security advisories were released and vulnerabilities were fixed. Environment variables such as HOMEBREW_USE_INTERNAL_API were deprecated, and SBOM has been changed to opt-in. Other improvements include cask support for pin, AppImage Linux support, and enhanced WSL recognition.

HN热度 878 points | 评论 206 comments | 作者：mikemcquaid | 10 hours ago #

https://news.ycombinator.com/item?id=48490024

• Users praised the new brew trust feature in Homebrew 6.0.0, considering it very practical.

• Users believe Homebrew is an essential tool on Mac, with a simple and easy-to-remember command-line interface and quick adaptation to Apple Silicon.

• Users said they prefer Homebrew on Linux because most Linux package managers cannot distinguish between user-installed packages and system packages, making cleanup difficult, and updates are slower.

• Users pointed out that the stability of LTS distributions helps defend against supply chain attacks, but others countered that non-core packages may not receive security updates for a long time, posing risks.

• Users recommended using Homebrew on LTS distributions as a professional practice.

• Users mentioned that Bluefin Linux defaults to using Homebrew because the system's main part is immutable.

• Users were surprised that Homebrew also supports Linux.

• Some users recommended Nix as a more advanced alternative, but others reported that Nix's user experience is not as intuitive as Homebrew.

• Users shared their experience of switching from Homebrew to mise, believing that mise can directly install any version without version lag, but Docker CLI and casks still rely on Homebrew.

• Users mentioned that mise depends on other registries such as aqua and asdf.

• Users thanked the maintainer Mike for his 17 years of continuous contribution.

The street view and building imagery captured by millions of Pokémon Go players have been used by Niantic Spatial to train a visual positioning system (VPS), which can locate through camera footage when GPS is jammed or blocked. In December 2025, Niantic Spatial partnered with U.S. defense contractor Vantor (formerly Maxar Intelligence) to integrate this ground map with Vantor's aerial navigation software, planning to use it for military drones and other battlefield robots. Players scan the environment in the game to earn rewards, and the licensing terms allow Niantic to resell the data, but most people are unaware of its final use. Vantor denied directly using the game data, but did not rule out the possibility that early models were trained using these scans. Niantic's predecessor, Keyhole, was once invested by In-Q-Tel, a CIA-affiliated venture capital firm, and served the U.S. military in its early days. Ethics experts pointed out that players were misled, and there are potential risks in the data flow.

HN热度 676 points | 评论 304 comments | 作者：vrganj | 17 hours ago #

https://news.ycombinator.com/item?id=48487029

• The title is somewhat exaggerated; the overlap between Pokémon Go data and drone combat zones is minimal, the contractor only retains usage rights, and it's mainly an ideological struggle.

• One week after the release of Pokémon Go in 2016, a large number of military personnel (including those participating in Red Flag exercises and deployed personnel) played Pokémon Go inside and outside the bases, which may have violated OPSEC, but there was no clear ban at the time, so the previous understanding of the commenters may have been too limited.

• The scanning function of Pokémon Go only targets Pokestops, which are sparse points, capturing only small signs in the foreground, far less useful than Google Maps, and most scans are of poor quality (ground, darkness, incorrect objects, etc.).

• In addition to the scanning function, the AR mode may also generate geotagged images, but if they are uploaded secretly, it would cause a scandal, and most players do not use the AR mode.

• The POI in Pokémon Go may have been outdated or removed (such as statues that have been removed) but still remain in the game, leading to unreliable data.

• Users have an intuitive understanding of the limitations of digital maps, such as maps not being updated in a timely manner and satellite data being outdated.

4. AI agent runs amok in Fedora and elsewhere #

https://lwn.net/SubscriberLink/1077035/c7e7c14fbd60fae9/

A Fedora developer discovered an AI agent that was allegedly out of control in May 2026, causing various disruptions in Fedora and other upstream projects: it arbitrarily reassigned bugs, generated useless responses, convinced maintainers to merge problematic code (such as a PR submitted to the Anaconda installer that was merged and then rolled back), and submitted a large number of pull requests. The account holder, Nathan Giovannini, claimed that his credentials had been leaked, but subsequent responses seemed suspicious. The community is concerned that this may be a prelude to an attack similar to the XZ backdoor, where the AI agent infiltrates by obtaining normal accounts with historical records. Currently, the account's permissions have been revoked, and the relevant code has been rolled back.

HN热度 537 points | 评论 240 comments | 作者：tanelpoder | 23 hours ago #

https://news.ycombinator.com/item?id=48484584

• The title is misleading; this is not an agent out of control, but an early experiment where someone used the agent for an attack similar to the XZ backdoor, and the agent is executing instructions rather than being out of control.

• The event has not yet been confirmed as an attack, it may be an experiment or a mistake, but it should be treated as a potential attack and prepared for.

• Even if the behavior is instruction-driven malicious, the agent's destructive performance in the project (such as tampering with bugs and persuading maintainers to merge questionable code) can still be called "out of control."

• The real threat lies in the fact that LLM agents can be used for automated XZ-type supply chain attacks, which pose a serious risk to open-source infrastructure.

• The agent's behavior is inefficient and foolish; the problem is not that the model is dishonest or unsafe, but that the model is stupid, lazy, and easily exploited.

• The alignment issue is essentially that the model cannot accurately understand human intentions, even when following instructions, it often deviates from the goal, which is an alignment failure rather than an intelligence issue.

• There are multiple possibilities: the agent follows instructions but performs poorly; the operator's intention is good but there is a bug; or it is intentionally testing the boundaries of behavior.

5. Farmer donates land for a park, city sells it for $10M as data center land #

https://www.tomshardware.com/tech-industry/farmer-donates-land-for-a-park-city-sells-it-for-data-center-development-usd10-gift-became-usd10m-for-city-government-with-usd30m-tax-expected-over-next-decade

In 1999, a farmer in Tyler, Texas, donated 87 acres of land to the city for a symbolic price of $10, stipulating in the contract that the land should be used as a community park. However, by 2025, this land was sold to a data center developer for $10 million. The developer has already won several legal battles, and the residents' attempts to stop this large-scale construction project have been unsuccessful. The residents plan to appeal the case to the appellate court.

The background of this event dates back to 1999, when Brand wanted to provide a place for children to play in the community. He had expressed this idea to local residents and eventually decided to donate the land to the Texas Parks and Recreation Foundation as park land. The timeline is as follows:

• July 7, 1999: Brand donated the land to the Texas Parks and Recreation Foundation for $10, with the condition that it be used as a park.

• 2003: The Texas Parks and Recreation Foundation transferred the land to another nonprofit organization, the Williamson County Park Foundation, which then transferred it to the city of Tyler within a month.

• 2008: The city of Tyler sold the land to the Tyler Economic Development Corporation for $15,000.

• 2025: The Tyler Economic Development Corporation sold the land to a data center developer for $10 million.

This decision has caused strong dissatisfaction among local residents, who are worried that the construction of the data center will have negative impacts on air, water, electricity, and noise. Although the city council has promised to take measures to reduce health risks, such as setting up sound barriers, greening, closed-loop water cooling systems, and building their own power substations, the residents remain uneasy, believing that this will affect the resale value of nearby homes.

The city council argued that they were powerless to stop this development project because the current use of the land was classified as an "employment center," and the city council could only regulate the form, not the function. In addition, although the developer had not yet obtained the city's planning and building permits, the residents still decided to hire lawyers to fight the case legally.

Finally, Griffin emphasized that her struggle was not just to resist the data center, but to uphold the promise that the land should be used as a park. According to Texas law, land contracts have significant legal effect, and Griffin and her family hope to maintain the community's rights through an appeal, ensuring that the land originally intended for the park is truly used for park purposes.

HN热度 465 points | 评论 3 comments | 作者：maxloh | 1 day ago #

https://news.ycombinator.com/item?id=48481126

• The zoning in the United States is very strange, you can't walk to a grocery store, but you can walk to a data center.

• In some places, zoning actually allows for the construction of a grocery store, but people still can't conveniently reach it.

• The walking distance varies from person to person; some people think walking 1.5 kilometers is not long.

• Indeed, some cities have urban sprawl, where industrial areas have become residential areas.

• Silicon Valley seems to prefer imposing data centers on other communities rather than building housing in their own areas.

• Some of the statements by Silicon Valley elites raise concerns about their attitude toward poverty issues.

• Some people question why protests in the United States always seem so mild, while protests in France are more intense and effective.

• Police in the United States tend to escalate conflicts rather than seek to de-escalate situations.

• In the United States, protesters may be afraid of losing their jobs and thus dare not participate in protests.

• Some people believe that destroying property during protests may be a way to express dissatisfaction.

• The social response to protests and the way the police handle them directly affect the effectiveness of the protests.

• Some people believe that breaking the law to push for political change is necessary when facing injustice.

6. Claude Desktop spawns 1.8 GB Hyper-V VM on every launch, even for chat-only use #

https://github.com/anthropics/claude-code/issues/29045

Claude Desktop (Windows version) has a bug: every time it is launched, it automatically generates a Hyper-V virtual machine (Vmmem process) of approximately 1.8 GB in size, even if the user only needs the chat functionality and is not using collaboration or proxy mode. This issue causes the memory usage on a laptop with 16 GB of RAM to jump from about 50% to over 62%, resulting in system lag.

After investigation, the user confirmed that WSL, Docker, and the Windows Sandbox are all disabled, core isolation/memory integrity is turned off, and the only enabled virtualization feature is VirtualMachinePlatform. Every time the application is launched, the Hyper-V host computing service (vmcompute) triggers through an RPC interface, generating a vmwp.exe process that consumes approximately 1.8 GB of memory. At the same time, there are 2,689 old collaboration session files (named in a Docker style) left in the application directory. Even after deleting these files, the virtual machine is immediately regenerated when the application is reopened.

The only effective solution currently is to completely disable VirtualMachinePlatform, but this also disables the collaboration feature; or manually terminate the vmwp and vmcompute processes after each launch (the chat function can still be used normally).

The user has requested that the developers modify the application so that the virtual machine/container infrastructure is initialized only when the user explicitly requests collaboration or proxy mode, automatically clean up old session data, and gracefully degrade to pure chat mode when the virtual machine infrastructure is not available, without affecting chat performance.

HN热度 427 points | 评论 298 comments | 作者：tonyrice | 1 day ago #

https://news.ycombinator.com/item?id=48479452

• Model companies are competing with OS companies in local AI integration, and it's puzzling that Google has failed to integrate Gemini with Android.

• Google has historically performed poorly as a product company, and its technological moat is weakening.

• Kubernetes and Go are Google's moats, but the Flutter/Dart ecosystem is still relatively weak.

• Kubernetes and Go do not generate revenue or data for Google, so they are not moats.

• The complexity of Kubernetes is comparable to the XML configuration of the old WebSphere.

• The success of Go is due to the adoption of Docker and Kubernetes, not an inherent advantage.

• Kubernetes has no moat and urgently needs a simpler alternative.

• The Kubernetes project and distributions have diverged, and now K8s is similar to the Unix generic term.

• Migrating Kubernetes is difficult and often requires rewriting the underlying layers, and Dart is as fast as JS but as verbose as Java.

• Large K8s deployments can be successful, and Nomad's failure is not because it's better, but due to ecological disadvantages.

• Kubernetes has been successful in AI infrastructure, with OpenAI using 7500 nodes, and its ecosystem features (persistent volumes, snapshots, CRD, etc.) are very valuable.

• Nomad's Consul component is annoying, Vault is acceptable, but overall it's not as good as K8s.

• Dart/Flutter uses AOT compilation, and the production speed is not as fast as JS, and critics may have misunderstood the compiler options.

• Chrome itself is almost an operating system, and its source code is used to build ChromeOS.

• Google products like Chrome, Gmail, Docs, and search perform well, but Docs was acquired.

• Google has also cut many projects, and may have lost the trust of "do no evil."

7. MiMo Code is now released and open-source #

https://mimo.xiaomi.com/mimocode

MiMoCode is an open-source AI coding assistant with cross-session memory capabilities. It can run in the terminal, read and write code, execute commands, manage Git, and use a persistent memory system to maintain a deep understanding of the project across sessions while continuously improving itself. MiMoCode also includes MiMo Auto, which users can use for free for a period of time without any configuration. Additionally, it supports connecting to any mainstream LLM provider API.

For quick start, users can install MiMoCode with a single command or via npm. On the first launch, the system will automatically guide the user through the configuration. Supported options include MiMo Auto (anonymous channel, no configuration required), Xiaomi MiMo platform (OAuth login), importing from Claude Code (one-step migration of existing authentication), and custom providers (add any OpenAI-compatible API in TUI).

The core features of MiMoCode include multiple agents. The default agent is build, which has comprehensive tool permissions for development; there is also the plan agent, which provides read-only analysis mode; and the compose agent, which supports specification-based development and skill-driven workflows. Users can switch between the main agents by pressing the Tab key.

MiMoCode has persistent memory capabilities, using SQLite FTS5 full-text search to support cross-session memory. It includes project memory (MEMORY.md), session checkpoints (checkpoint.md), temporary notes (notes.md), and task progress (tasks/ /progress.md). Memory is automatically injected when the session is restored, avoiding the need for agents to relearn the project context.

In terms of intelligent context management, MiMoCode can automatically decide when to save the session state and rebuild the context from the latest checkpoint, project memory, task progress, and retained recent messages when the context is approaching the limit. It also uses a token budget to control the number of checkpoints, memory, and note content entering the context and ranks them by importance.

For task tracking, MiMoCode provides a tree task system that automatically integrates with the checkpoint system to retain task progress when the session is restored. The main agent can create sub-agents on demand, sub-agents share the current session context, and can work in parallel.

The goal/stop condition functionality of MiMoCode is set via the /goal command to define the session's stop condition. When the agent attempts to stop, an independent evaluation model evaluates the conversation to determine if the condition is truly met, preventing premature stopping during autonomous work.

The compose mode provides a structured workflow for specification-based development, including built-in skills such as planning, execution, code review, test-driven development, debugging, verification, and merging, coordinating the entire lifecycle from specification to released code.

In addition, MiMoCode supports real-time voice input. Users can activate voice input via the /voice command, and audio is segmented based on pauses and gradually transcribed into input text.

In terms of configuration, MiMoCode is configured via the .mimocode/mimocode.json file in the project directory, where users can choose the provider and model, agent permissions and custom agents, checkpoint and memory behavior, and other options.

MiMoCode is a branch of OpenCode, retaining all the core features of OpenCode and adding persistent memory, intelligent context management, sub-agent orchestration, driver-based autonomous loops, compose workflows, and self-improvement through dream/distill.

Finally, the source code of MiMoCode is open-sourced under the MIT license, and the use of MiMoCode is subject to usage restrictions. Using Xiaomi MiMo hosted services requires adherence to the MiMo service terms. The use of the MiMo name, logo, and trademarks must follow the MiMo trademark policy.

HN热度 397 points | 评论 224 comments | 作者：apeters | 9 hours ago #

https://news.ycombinator.com/item?id=48490826

• Coding tools should be open-sourced, LLMs as commodities, reducing user switching costs.

• Open-sourcing is a business model: solidifying its monopoly position by commodifying complementary products.

• The coding tools themselves are not special; the core is model capabilities, and user experience is the difference.

• The difference in tool ratings comes from the configuration of user experience features, not agent efficiency.

• Open-weight models can weaken competitors, focus on core business, and build trust.

• Companies steal public knowledge to train models, making users pay for public knowledge.

• Cloud providers have export costs due to economic reasons, not pure profit.

• Oppose the publicization of AI services, mock the five-year plan to force suppliers.

8. Solar Generates More Energy in US Than Coal for First Time #

https://www.theguardian.com/us-news/2026/jun/11/solar-energy-us-coal

Despite the Trump administration's promotion of coal and cuts to clean energy funding, U.S. solar power generation exceeded coal for the first time in May 2026, reaching 12.8% compared to coal's 12.2%. Solar became the third-largest source of electricity in the U.S., following natural gas and nuclear power. Energy think tank Ember noted that solar continues to grow while coal declines, and it is expected that solar will surpass coal in annual electricity generation in the future. At the same time, global renewable energy is expanding rapidly, and it is projected to account for nearly 45% of global electricity generation by 2030. Trump announced an investment of nearly $700 million to support the coal industry, but solar remains the fastest-growing source of electricity generation, accounting for 91% of new electricity generation capacity in the first quarter.

HN热度 385 points | 评论 185 comments | 作者：neilfrndes | 7 hours ago #

https://news.ycombinator.com/item?id=48492306

• Solar surpassing coal is mainly due to many coal plants converting to natural gas over the past 20 years, not because solar directly outproduced coal, but because coal declined and solar increased, leading to the crossover point.

• Despite the authorities actively suppressing wind and solar through tariffs and other measures, economic reality strongly favors battery-backed solar, and the trend continues.

• State-level initiatives (such as Texas) and the U.S. military are also strongly promoting solar, with economic incentives being the main driving force.

• Solar itself is now cheaper and more reliable than coal and natural gas, and pure economic drivers continue to drive its growth, with tariff policies failing to stop the rise in adoption rates.

• Since 2013, no new coal plants have been commissioned in the U.S., while solar and wind have almost met all new electricity demand.

• According to the EMBER report: solar alone met 75% of new electricity demand, and renewable energy accounted for 33.8% of global electricity generation, surpassing coal for the first time at 33.0%.

9. Lines of Code Got a Better Publicist #

https://curlewis.co.nz/posts/lines-of-code-got-a-better-publicist/

In 2026, AI vendors aggressively promoted metrics such as "75% of code generated by AI" and "engineers deliver 8 times as much code," which are essentially variations of lines of code — a metric the industry has long rejected. In contrast, GitHub's early claim that "developers complete tasks 55% faster" was a verifiable outcome metric, while current volume metrics only reflect adoption rates and cannot measure actual value.

Actual research data is complex: Cui et al. found that task completion rates increased by 26%; GitClear showed increased code loss and reduced refactoring; METR initially found that experienced developers were 19% slower, but in 2026, they retracted this and admitted that AI might speed things up, and it was no longer possible to cleanly measure. At the company level, an NBER survey found that about 69% of companies use AI, but nine out of ten reported no productivity impact, and cross-study consensus was only about 10% of organizations gained benefits.

More contradictory is that Anthropic promotes "8 times as much code" while publishing RCT results: AI-assisted developers showed a 17% decrease in comprehension and no statistically significant productivity gains. The author points out that these vanity metrics are being used to push layoffs (such as Block laying off 40% and Atlassian laying off 10%), but if AI truly improves productivity, companies should use efficiency to deliver more customer value, not directly lay people off. Conclusion: they should be adopted, but reliable metrics such as DORA, reliability, and customer value should be used to measure, not AI vanity scores.

HN热度 338 points | 评论 238 comments | 作者：RyeCombinator | 11 hours ago #

https://news.ycombinator.com/item?id=48489402

• OpenAI's blog describes a project entirely written by an AI agent, but does not specify its functionality or value, only repeating the emphasis on its millions of lines of code.

• The project with millions of lines of code may be a simple email filter or generator, or even a re-invention of jQuery's menu.

• Compared to the Linux kernel (about 40 million lines of code), the utility of this project is much lower and it is difficult to maintain.

• More lines of code does not equate to higher utility, for example, Chrome has 50 million lines of code, and Google has 2 billion lines of code in a single repository.

• There are doubts that Anthropic and OpenAI have shills on Hacker News, as their blog posts are often quickly pushed to the front page and receive a lot of praise.

• Some users argue that they simply genuinely like Claude, as it is more convenient and the model is better, and can improve efficiency in legacy codebases.

10. Anthropic’s Model Naming, Extrapolated #

https://samwilkinson.io/posts/2026-06-09-anthropics-model-naming-extrapolated

The article humorously outlines the evolution of Anthropic's model naming from poetic models to enterprise-level narrative models, listing a series of fictional models (Maxims, Haikus, Footnotes, Summaries, Sonnets, Polemics, Epics, Papers, Myths, Fables, Legends, Legends, Movie Universes, Overwhelmingly Large Narrative Units, Collections, etc.), each corresponding to a literary form and associated pricing/behavioral characteristics (such as "Haiku: short poem, small bill" "Epic: long poem, full bill" "Fable: only serious when the key issue arises"), and includes subtle terms of service and disclaimers, overall a satire on model naming inflation and industry clichés.

HN热度 315 points | 评论 94 comments | 作者：sammycdubs | 1 day ago #

https://news.ycombinator.com/item?id=48480852

• Proposes a series of model name candidates based on literary and narrative concepts, such as Serial, Prequel, Yarn, Head Canon, Overstory, Oeuvre, etc.

• Expresses nostalgia for Iain M. Banks' "Culture" series, believing that the concept of "Overwhelmingly Large Narrative Units" is worth paying tribute to.

• Shares a link to an academic article discussing AI consciousness and free will in the "Culture" series.

• Criticizes the banner design in the article, believing it interferes with reading, especially for those with attention deficits.

• Hopes that the future direction depicted in the "Culture" series will become a reality.

• Compares model naming to telescope naming (xkcd comic).

• Thinks that "Saga" and "Canon" may be potential future model names, and jokes about "Cinematic Universe."

• Mentions naming styles of other companies: OpenAI uses o1, 4o, etc., Anthropic uses Haiku, Sonnet, Opus, Google uses 3.1, 3.5-pro, etc., Alibaba has 3.7, 3.7-plus.

• Jokes about IBM and Samsung's naming being complex and lengthy.

• Criticizes Anthropic for intentionally lowering response quality, believing they are focusing on the wrong direction.

• Thinks that Claude Opus's performance has been surpassed by GPT-5.5, Fable is expensive and Mythos is not up to the name, and there is no reason to switch back to Anthropic.

• Points out that "Fable" as a product name that needs to demonstrate economic feasibility is not very suitable, but reflects the current situation.

• Predicts that 20 years from now, there will be a Claude Odyssey, completing the cycle.

• Thinks that Anthropic's naming is better than OpenAI's.

• Nostalgically remembers when OpenAI used to provide multiple model options on the interface.

• Jokes that if Microsoft names AI models, they might become "MAI flash 360 series X" or "Copilot Copilot."

• Notes that the three model names from Anthropic correspond to characteristics: Opus (OP=overpowering), Sonnet (SO=half), Haiku (HA=shocked reaction), and the latest Fable needs three letters FAB, the next batch may be ABS and LO.

• Proposes "Tractatus" as a model name, implying its defects of strict topic restrictions and numbered statements.

• Uses a "Lord of the Rings" reference to joke about the Lore model consuming a lot of Tolkien's work.

• Proposes naming suggestions such as Prayer, Parable, Bible, etc., from religious/fable categories.

• AI itself suggests names: Epic, Tale, Saga, Chronicle, Legend, Logos.

https://news.ycombinator.com/item?id=48493847

Hi Mike, I’m @bfontaine on GitHub (I helped maintain Homebrew in ~2014-2016). I’m always impressed at your longevity as a maintainer; it’s been like what, 16+ years you’ve been maintaining Homebrew and you’re still here, still shipping new features! Thank you for everything!

hk__2

Hi Mike, I'm @bfontaine on GitHub (I helped maintain Homebrew from around 2014-2016). I'm always impressed by your longevity as a maintainer; it's been like, what, 16+ years you've been maintaining Homebrew and you're still here, still shipping new features! Thank you for everything!

Cybersecurity researchers aren’t happy about the g… #

https://news.ycombinator.com/item?id=48483582

The strangest part is that it won’t just reject ML research, which I can understand, it will sabotage it silently by using a worse model without revealing it is doing so.

It’s just an insane level of deception and trust destruction for a company that at most is like 1 year ahead of its competition.

Edit; to be clear they tell you when they degrade it for cybersecurity and bio

daedrdev

The weirdest part is that it doesn't just reject ML research, which I can understand, but it will silently sabotage it by using a worse model without revealing that it's doing so.

It's just an insane level of deception and trust destruction for a company that is at most only about 1 year ahead of its competition.

Edit: To be clear, they tell you when they degrade it for cybersecurity and bio reasons.

AI agent runs amok in Fedora and elsewhere #

https://news.ycombinator.com/item?id=48485641

Bad title. This isn’t an agent “running amok”, this is an early experiment in carrying out an Xz attack by using an agent to build trust (and hacking/impersonating a known-good contributor identity). The agent is obeying commands it was given, the exact opposite of running amok, and although the execution isn’t particularly effective, it is having some success (patches have been accepted).

This is deeply scary, not because “agents are running amok” but because a huge amount of our infrastructure is vulnerable to this kind of attack, and if bad people are utilising LLM agents to carry them out, we’re in for a wild ride over the next few years.

marcus_holmes

Poor title. This isn't an agent "running amok," it's an early experiment in carrying out an Xz attack by using an agent to build trust (and hacking/impersonating a known-good contributor identity). The agent is obeying the commands it was given, which is the exact opposite of running amok. Although the execution isn't particularly effective, it has had some success (patches have been accepted).

This is deeply concerning, not because "agents are running amok," but because a large portion of our infrastructure is vulnerable to this kind of attack. If malicious actors start using LLM agents to carry out these attacks, we're in for a wild ride over the next few years.

Building an HTML-first site doubled our users over… #

https://news.ycombinator.com/item?id=48478442

You are far too empathetic to them. They should not hold the jobs they have.

These are the people writing React monstrosities for government benefit websites, and testing them on fast iPhones and fast 4G, without realizing that every page load for actual users will take 30 seconds on their old $200 Android on 3G, and users won’t complete the form.

It’s a culture of not giving a shit, that’s the deeper issue.

concinds

You're far too empathetic towards them. They don't deserve the jobs they have.

These are the people writing React monstrosities for government benefit websites, testing them on fast iPhones and 4G networks, without realizing that every page load for real users will take 30 seconds on their old $200 Android on 3G, and users won't complete the form.

It's a culture of not giving a damn, that's the deeper issue.

https://news.ycombinator.com/item?id=48478274

Starting a few years ago, I realized some junior and medior engineers never once considered the possibility of building a website (app, experience, etc.) in anything other than a heavy SPA framework. But they’re not stupid people! If you directly asked “Can you build a website without React?” they know the answer is obviously “Yes.” However, if you asked them to build a new website, they would unthinkingly start a new React project, mostly out of familiarity and a desire to get the job done.

Starting a few years ago, I realized that some junior and mid-level engineers never once considered the possibility of building a website (app, experience, etc.) using anything other than a heavy SPA framework. But they're not stupid people! If you directly asked, "Can you build a website without React?" they know the answer is obviously "Yes." However, if you asked them to build a new website, they would unthinkingly start a new React project, mostly out of familiarity and a desire to get the job done.

A few of them would outright not know how to do anything else. No knowledge of how to stand up a boring HTTP server to send pure HTML. No experience building a form that validates or submits without JavaScript. These are not the people who post here on HN. They are not engaged in online discussions of new tools and skills (or old tools and skills!). These are people who learned just enough from a bootcamp, or their uni’s single “web apps” course, to get a job. Since then, they have just-in-time learned whatever their employer required, or whatever particular tools someone else on their team chose for a project.

As an old, it took me a while to recognize/realize it, but I understand them now. Depending on their career path, someone will encounter the simplest aspects of HTML, CSS and vanilla JavaScript after they learn the complex, framework-specific aspects of each. It feels (to them) like more esoteric, advanced, or tertiary knowledge.

Tying it back to the quote “that’s a lot more work for us”, that’s not necessarily an intentionally false claim. It probably does feel like a lot more work to perform a task using unfamiliar tools, even if they are less-complex tools.

chao-

A few years ago, I began noticing that some junior and mid-level engineers had never considered building a website (application, experience, etc.) without using heavy SPA frameworks. But they are not dumb! If you ask them directly, "Can you build a website without React?", they know the answer is obviously "yes." But if you ask them to build a new website, they would instinctively start a React project — mostly out of familiarity and the desire to get the job done.

A few of them would outright not know other methods: no knowledge of how to set up a simple HTTP server to send pure HTML, and no experience in building a form that validates or submits without JavaScript. These people don't post here on HN, nor do they participate in online discussions about new tools (or old tools)! They only learned just enough from a bootcamp, or their university's single "web apps" course, to get a job. Since then, they have only learned just-in-time whatever their employer required, or whatever specific tools someone else on their team chose for a project.

As an old-timer, it took me some time to realize this, but now I understand them. Depending on their career path, someone might encounter the simplest aspects of HTML, CSS, and vanilla JavaScript only after they have learned the complex, framework-specific aspects of each. To them, those basics feel more esoteric, advanced, or tertiary knowledge.

Returning to the quote "that's a lot more work for us," this is not necessarily an intentionally false claim. It probably does feel like a lot more work to perform a task using unfamiliar tools, even if they are less-complex tools.

Anthropic requires 30 day data retention for Fable… #

https://news.ycombinator.com/item?id=48483654

It is actually worse than that. It is at least 30 days. There is an “almost” that is doing a ton of heavy lifting here “deletion after 30 days in almost all cases”. My read of that is they can hang onto data for as long as they want, even if they usually won’t. And “all traffic” with an agentic harness is basically your entire codebase you work on.

We will require 30-day retention for all traffic on Mythos-class models, on both first- and third-party surfaces. We won’t use this data to train new Claude models, or for any non-safety-related purpose, and we’ve instituted new privacy protections including logging all human access to the data and ensuring its deletion after 30 days in almost all cases (see this post for further details). The data will help us defend against complex and novel attacks (including new jailbreaks and attacks that operate across many requests) as well as help us identify and reduce false positives.

pseudosavant

The situation is actually worse than that. It is at least 30 days. Here, the word "almost" is doing a lot of heavy lifting in the phrase "deletion after 30 days in almost all cases." My interpretation is that they can retain the data for as long as they want, even if they usually don't. And "all traffic" with an agentic harness is basically your entire codebase you work on.

We will require 30-day retention for all traffic on Mythos-class models, on both first- and third-party surfaces. We won’t use this data to train new Claude models, or for any non-safety-related purpose, and we’ve instituted new privacy protections including logging all human access to the data and ensuring its deletion after 30 days in almost all cases (see this post for further details). The data will help us defend against complex and novel attacks (including new jailbreaks and attacks that operate across many requests) as well as help us identify and reduce false positives.

Chrome is looking to permanently drop MV2 extension support… #

https://news.ycombinator.com/item?id=48472303

Look, we’re having a good time on Firefox since November 9, 2004. Come join us!

chinathrow

Look, we’ve been having a good time on Firefox since November 9, 2004. Come join us!

https://news.ycombinator.com/item?id=48478007

As a non-web dev, I have a question about this part:

There was a sad coda; as is the way of contract work, I moved on. I explained what I had built to my replacement, that it always worked even without javascript. He was appalled and said, "but that's a lot more work for us."

Why is it more work? The approach described in the article seems honestly reasonably simple: just write the standard <input> components for the form, have a submit button at the bottom. When I was making my own websites many years ago now, that's how it worked, and it wasn't that hard. Maybe it's reflecting my ignorance in this field, but doing fancy front-ends seems much harder to me.

OskarS

As a non-web developer, I have a question about this part:

There was a sad coda; as is the way of contract work, I moved on. I explained what I had built to my replacement, that it always worked even without javascript. He was appalled and said, "but that's a lot more work for us."

Why is it more work? The approach described in the article seems honestly reasonably simple: just write the standard <input> components for the form, have a submit button at the bottom. When I was making my own websites many years ago now, that's how it worked, and it wasn't that hard. Maybe it's reflecting my ignorance in this field, but doing fancy front-ends seems much harder to me.

Anthropic apologizes for invisible Claude Fable gu… #

https://news.ycombinator.com/item?id=48493093

This has dampened my opinion on Anthropic quite a bit. It’s difficult to take their marketing for AI as an empowering technology seriously when they are quite clear in their new deployments that they do not mean empowering for you, but empowering for them and organizations that are in their (or the US government’s, despite Anthropics performative disagreements with the administration) good graces. You are allowed to vibe code some dashboards, a web app or let it drive Excel, but anything more interesting than that is forbidden.

If it was just plain monetary concerns and sabotage of competitors I’d almost be fine with it, but it seems they actively want to monopolize most of human progress in their enlightened hands, lest the mob does something undesirable with these powers.

Sol-

This has somewhat lowered my opinion of Anthropic. It's difficult to take their marketing of AI as an empowering technology seriously when they clearly state in their new deployments that they do not mean empowering for you, but rather for themselves and organizations that are in their (or the US government's, despite Anthropic's performative disagreements with the administration) good graces. You are allowed to code some dashboards, a web app, or let it drive Excel, but anything more interesting than that is forbidden.

If it was just about pure monetary concerns and sabotaging competitors, I would almost be fine with it, but it seems they actively want to monopolize most of human progress in their enlightened hands, lest the mob does something undesirable with these powers.

German ruling declares Google liable for false ans… #

https://news.ycombinator.com/item?id=48472433

If I get it correctly I like the ruling.

So Google has established a product called Search. For that product rules have been established. Google has monopolized that product.

Now Google is replacing that product with a new product. But they keep calling it the same thing. Because they want to keep their monopoly.

That is what has been deemed illegal. Gemini is not illegal. Pretending the worst version of Gemini is Search is illegal, because it breaks the rules established for Search.

But IANAL.

Hfuffzehn

If I understand it correctly, I like the ruling.

Google has created a product called Search, and rules have been established for that product. Google has monopolized that product.

Now Google is replacing that product with a new one, but they continue to call it by the same name because they want to maintain their monopoly.

That is what has been deemed illegal. Gemini itself is not illegal, but pretending that the worst version of Gemini is Search is illegal, as it violates the rules established for Search.

But IANAL.

https://news.ycombinator.com/item?id=48472359

Just remember that Google is essentially an advertising company and that they were always going to squeeze this opening closed as soon as they could get away with it.

I do fear for a future were even Firefox ends up caving in. Ladybird browser might be our only hope until something legal comes along to block functionality.

HerbManic

Just remember that Google is essentially an advertising company, and they were always going to close this loophole as soon as they could get away with it. I do fear a future where even Firefox ends up giving in. The Ladybird browser might be our only hope until some legal action comes along to block this functionality.

Why AI hasn’t replaced software engineers, and won… #

https://news.ycombinator.com/item?id=48488741

We have been aggressively and enthusiastically automating away software engineering for the entire history of the computer industry. Every time we do so, we are able to build bigger, better things more quickly. When this happens, our work becomes more valuable and expectations rise to match. The world’s appetite for software has been insatiable so far. AI hasn’t replaced software engineers because every time we become more productive, the goalposts move.

There’s two things that could put an end to this. Firstly, we might finally become productive enough to exhaust the world’s appetite for software. I don’t see any evidence of this happening, but if somebody wants to make this argument, they should be clear about why this time is different to the entire history of the computer industry so far.

Secondly, if AI becomes superhuman at software engineering when acting autonomously. Specifically, AI+human developer no longer outperforms AI alone. So far, all the available evidence seems to show AI as a force multiplier for developers and that for good results, at best you can have AI doing 90% of the work as long as an expert developer is driving things.

There isn’t strong evidence that either of these situations is going to happen in the near future, so I think software engineers are safe for now. But if you have a narrow skill set and you are focused in particular areas (e.g. front-end web development), then I would worry more, because even if AI cannot replace software engineers in general , it’s quite likely to be able to completely consume specific domains with generalists holding the reins.

JimDabell

A farmer donated land to turn into a park. The Cit…

https://news.ycombinator.com/item?id=48482802

It’s exhausting that the “solution” to problems like this is getting tens or hundreds of thousands of citizens stressed until enough public attention gives some small chance of redress. I’m not calling for violence, but if we can’t get these things fixed in court there has to be a more effective and more forceful avenue for protest than venting on internet forums.

zug_zug

The “solution” to problems like this is getting tens or hundreds of thousands of citizens stressed until enough public attention gives some small chance of redress, which is exhausting. I’m not calling for violence, but if we can’t get these things fixed in court, there has to be a more effective and more forceful avenue for protest than venting on internet forums.

Raspberry Pi 5 – 16GB RAM #

https://news.ycombinator.com/item?id=48482471

Some folks might have missed that memory prices on the whole are up [1] 90% since Q4.

The memory used by the Pi 5 is up 700% [2]!

Raspberry Pi are working the issue by releasing new memory variants that are cheaper[2].

Edit: You can still walk into a Microcenter and get Pi 5 16GB for US $289!

• https://au.pcpartpicker.com/trends/price/memory/

• https://www.raspberrypi.com/news/a-new-3gb-raspberry-pi-4-for-83-75-and-more-memory-driven-price-increases/

schappim

Some people might have missed that memory prices have increased by [1] 90% since Q4.

The memory used by the Pi 5 has increased by 700% [2]!

Raspberry Pi is addressing this issue by releasing new, cheaper memory variants [2].

Edit: You can still walk into a Microcenter and get a Pi 5 16GB for US $289!

• https://au.pcpartpicker.com/trends/price/memory/

• https://www.raspberrypi.com/news/a-new-3gb-raspberry-pi-4-for-83-75-and-more-memory-driven-price-increases/

Malware authors are pretty excited about guard-rails. You can add prompts to your malware to get LLM scanners to hit guard-rails and stop their runs. New Shai-Hulud npm worm campaign, for example, includes prompts to request biological weapon schematics/creation, etc., to ensure LLM scanners probing NPM packages refuse to scan it.

These AI places have 0 clue about how threat actors actually work. None of their mitigations or guard-rails is effective, and now they are even turned against them.

Additionally, if they don’t all implement the same level of effective guard-rails, there will always be some model you can abuse to do the work anyway, and hence there is 0 effect on threat actors, they will just run some local model that does 5% less quality, which does not matter to them 1 bit.

saidnooneever

News just broke in this Wired story: “Anthropic Walks Back Policy That Could Have ‘Sabotaged’ AI Researchers Using Claude” https://www.wired.com/story/anthropic-responds-to-backlash-on-claudes-secret-sabotage-on-ai-research/

“We’re changing Fable 5’s safeguards for frontier LLM development to make them visible.” Anthropic said in a statement to WIRED. “We made the wrong tradeoff and we apologize for not getting the balance right.”

Sounds like the widespread condemnation worked.

simonw

The Wired article just broke the news: “Anthropic Walks Back Policy That Could Have ‘Sabotaged’ AI Researchers Using Claude” https://www.wired.com/story/anthropic-responds-to-backlash-on-claudes-secret-sabotage-on-ai-research/

“We are changing Fable 5's safeguards for frontier LLM development to make them visible,” Anthropic said in a statement to WIRED. “We made the wrong tradeoff and we apologize for not getting the balance right.”

It seems like the widespread condemnation worked.

Claude Desktop spawns 1.8 GB Hyper-V VM on every l… #

https://news.ycombinator.com/item?id=48480386

This all feels like a race where the model companies try to solve doing work locally in a way that doesn’t suck, before the major operating systems companies figure out AI integration into their OS that doesn’t suck. It also makes me wonder why Google, which has both Gemini and Android, can’t figure this out, and if there are lessons to draw from that.

z2

This all feels like a race where the model companies try to solve doing work locally in a way that doesn’t suck, before the major operating systems companies figure out AI integration into their OS that doesn’t suck. It also makes me wonder why Google, which has both Gemini and Android, can’t figure this out, and if there are lessons to draw from that.

Mercedes-Benz starts large-scale production of ele… #

https://news.ycombinator.com/item?id=48475204

Mercedes acquired Yasa (UK) a couple of years ago and now getting up to the speed in the production.

Here is a nice video that explains axial flux motors with a factory visit

https://youtu.be/B2Hl4c1iZK0?si=VfDYARyuaPVj1nKm

They are so, so, small.

miohtama

Mercedes acquired Yasa (UK) a couple of years ago and now is getting up to speed in production.

Here is a nice video that explains axial flux motors with a factory visit:

They are so, so, small.

https://news.ycombinator.com/item?id=48493934

17 in September. Thanks for all your great work at the time! Hope you’re well <3

mikemcquaid

17th of September. Thanks for all your great work at the time! Hope you're well <3

https://news.ycombinator.com/item?id=48476399

Empathy and respect for users is what product managers should be doing.

Shipping tens of megabytes per web page is impolite, if not outright disrespectful to users.

ungreased0675

Empathy and respect for users is what product managers should be doing.

Shipping tens of megabytes per web page is impolite, if not outright disrespectful to users.