v2026.09.01

Tuesday, September 1, 2026

partial

The Marías

Cariño

0:004:36

Weather

Bloomington

74°

Clear

H 95°L 74°

Feels like: 77°
Humidity: 79%
Cloud: 0%
Rain: 1%

Today

95°

74°

Wed

97°

74°

Thu

100°

71°

Fri

100°

76°

Sat

98°

72°

Sun

91°

68°

Mon

95°

65°

World Clocks

San Francisco

—

New York

—

London

—

Berlin

—

Dubai

—

Mumbai

—

Shanghai

—

Sydney

—

Markets

Tesla led the session with a 5.51% gain alongside Qualcomm and crude oil, but major tech stocks traded in a split tape. Semiconductor and hardware names notched modest gains, while platform heavyweights like Amazon and Alphabet pulled down software.

Version v2026.09.01 - Daily Briefing

The Debate

If the British Parliament had accepted the Conciliation Plan of 1775 and retained the American colonies, would chattel slavery in North America have ended sooner than 1865?

Understanding this fork reveals whether institutional imperial governance or regional economic incentives serve as the primary driver of major moral realignments in modern history.

Faster Emancipation

If Parliament had conciliated with the American colonies in 1775, North America would have remained inside the British Imperial legal apparatus. The British Empire formally banned the slave trade in 1807 through the Abolition of the Slave Trade Act and ended colonial slavery entirely under the Slavery Abolition Act of 1833. Without a sovereign, slaveholding republic like the United States protecting the institution through constitutional compromises and southern political dominance, American planters would have lacked the geopolitical independence required to resist imperial legislation. Parliament’s willingness to compensate slaveholders, as done in Jamaica and Barbados, would have dismantled the plantation economy three decades before the American Civil War.

Slavery Entrenched

Retaining the Thirteen Colonies would have radically altered Parliament's internal political composition. In 1833, the West India Lobby held only a small minority of parliamentary seats, making imperial abolition politically feasible in London. Adding dozens of wealthy North American slaveholding representatives to Parliament would have created an unstoppable pro-slavery voting bloc. American planters generated far more cotton and tobacco wealth than Caribbean sugar estates. Faced with direct parliamentary representation from Virginia or South Carolina, Britain’s legislative majority against slavery would have evaporated. Far from accelerating emancipation, imperial integration would have transformed London into the political fortress of international slavery, delaying emancipation past 1865.

cruxWhether North American planter elites in Parliament could have formed a voting coalition powerful enough to block the 1833 Slavery Abolition Act.

Does constitutional judicial review derive its legitimacy from popular sovereignty, or does it inherently conflict with democratic legitimacy?

The answer determines whether constitutional safeguards against legislative majorities strengthen civic self-rule or undermine political equality.

Limit On Sovereignty

Democracy fundamentally requires that public policy reflects the ongoing political will of living citizens through elected representatives. When unelected judges invalidate legislation based on constitutional interpretations, they substitute judicial moral reasoning for popular majoritarian consent. Even if a constitution was originally ratified by a popular vote, holding current generations hostage to eighteenth-century ratifiers violates democratic self-governance. Judicial review creates an irreducible counter-majoritarian difficulty: it transforms legal arguments into supreme political decisions, stripping ordinary citizens of the power to decide crucial public questions. True democratic legitimacy rests in the legislature, where citizens bargain as political equals, not in a court enforcing anti-majoritarian vetoes.

Popular Sovereignty

Constitutional judicial review is not an anti-democratic constraint, but the essential mechanism by which a sovereign people protects its democratic self-binding commitments. Majorities in ordinary legislatures frequently face short-term electoral pressures that incentivize curtailing voting rights, political speech, and procedural fairness for minorities. By holding statute law to supermajoritarian constitutional standards like the born free and equal clauses of 1780, courts enforce the fundamental rules that make majoritarian decision-making legitimate in the first place. Judicial review ensures that current political majorities cannot alter the structural rules of democracy to entrench themselves. The judiciary acts as the guardian of the sovereign public's long-term democratic contract.

cruxWhether democratic legitimacy requires unconstrained majoritarian legislative power or adherence to pre-established structural ground rules.

Is human social order fundamentally maintained through internalized shared norms or through institutional coercion?

How we answer this defines whether building functional societies requires cultivating moral consensus or engineering effective coercive incentives.

Internalized Norms

Anthropological field studies across egalitarian foraging bands, non-state agricultural villages, and modern urban communities show that social order persists primarily through shared moral frameworks, reputational incentives, and ritualized norms. Formal coercive apparatuses like police forces and written penal codes are recent historical inventions that operate only at the fragile margins of society. No state authority possesses the surveillance capacity to police millions of daily human interactions. Instead, individuals self-police because they internalize moral concepts, social expectations, and mutual obligations. Institutional sanctions only function effectively when the overwhelming majority already views compliance as normative. Coercion is an emergency backup system for order, not its primary foundation.

Institutional Coercion

Social order is structurally defined and stabilized by monopolized violence and institutional enforcement, not warm-hearted consensus. Without formal mechanisms to penalize free-riders and defectors, informal norms collapse rapidly under population growth, social diversity, or resource scarcity. History demonstrates that whenever centralized enforcement mechanisms break down—whether in ancient urban collapses or modern civil war conflicts—shared cultural norms prove incapable of preventing opportunism and predation. Reputation and social shame only deter non-conformity in small-scale kinship groups; large-scale human cooperation depends on legal codes, tax collection, and the credible threat of physical force by recognized authorities to guarantee compliance across strangers.

cruxWhether large-scale social cooperation among non-relatives can remain stable over generations without centralized coercive enforcement.

From the Labs

Google Research/lab

Google Research releases TimesFM-3 for zero-shot multivariate forecasting

Google Research introduced TimesFM-3, a 330-million-parameter time-series foundation model natively pre-trained on over 1 trillion real-world and synthetic time points. Unlike earlier univariate iterations, TimesFM-3 supports multivariate forecasting in a single forward pass without task-specific fine-tuning. The model uses a decoder-only transformer architecture, grouping contiguous data points into patches of 32 time steps with per-series normalization. For past and future covariates, it uses a lookahead strategy by concatenating current and future patches into single tokens. The main transformer stack processes these tokens using a 2D grid that alternates between temporal self-attention across time and cross-series self-attention across features. Using Contiguous Patch Masking, it fills all future horizon patches simultaneously, outputting nine prediction quantiles from the 10th to 90th percentiles.

Multivariate forecasting is where practical time-series work actually lives. Real-world tasks like retail demand planning rely on coevolving factors such as product sales, foot traffic, and upcoming promotion schedules. Older patch-by-patch models generated forecasts sequentially, which created computational latency, accumulated errors, and spiked compute costs. By masking the future horizon and generating all predictions in one pass, TimesFM-3 eliminates iterative loops while incorporating known future events like planned holiday promotions. For instance, feeding a scheduled promotion calendar into the model lets it anticipate a roughly 20 percent sales lift on promotion days. Evaluated across public benchmarks including Gift-Eval and FEV-Bench, the architecture combines cross-series context with uncertainty estimation to make real-world operations significantly more actionable.

Google Research

Alteon/startup

Alteon raises $2.5 million to build long-endurance wind-powered aircraft

Bengaluru-based startup Alteon raised $2.5 million in a pre-seed round led by solo investor Lachy Groom, with participation from Together Fund. Founded in 2025 by 20-year-old Samay Sanghvi, Alteon is developing small autonomous aircraft designed to stay airborne over the ocean for up to a year without landing. The technology relies on dynamic soaring, an atmospheric maneuver modeled on albatross flight that extracts kinetic energy from wind shear above ocean surfaces. In a recent flight test over the Bay of Bengal, a prototype with a roughly three-meter wingspan completed seven autonomous O-shaped cycles at over 62 miles per hour within one meter of the ocean surface. Operating out of a 10,000-square-foot facility, the 20-person team builds four to five aircraft weekly.

If successful, ultra-long-endurance aircraft could transform maritime surveillance by offering continuous real-time monitoring for governments without expensive refueling logistics. Alteon's roadmap involves converting excess flight energy into electricity using propellers as turbines, eventually recharging onboard batteries. However, major technical hurdles remain before reaching commercial viability. Alteon has not yet achieved energy-neutral dynamic soaring, where propulsion is turned off entirely and flight is sustained purely by wind shear. Aerospace experts note that operating a drone just meters above breaking waves requires negotiating unpredictable local turbulence, sea spray, rain, and shifting light conditions. While the fundamental physics of dynamic soaring is proven, sustaining continuous flight in messy real-world marine environments is an unproven feat that demands extensive continuous flight testing.

TechCrunch

1789 Capital/funding

Polymarket raises $300 million from Donald Trump Jr.'s 1789 Capital

Prediction market Polymarket raised $300 million from 1789 Capital as part of a larger funding round totaling roughly $1 billion. 1789 Capital, an investment fund where Donald Trump Jr. serves as a partner, previously put $200 million into the platform. The firm is known for backing unconventional projects, including the tech-backed steroid Olympics known as the Enhanced Games. The massive cash injection arrives as prediction markets face escalating legal friction with state regulators across the country. At least 20 states are currently pursuing litigation against prediction sites over sports wagers, while a coalition of 44 state attorneys general signed a letter asserting that federal regulators lack authority over sports betting on these platforms.

The investment highlights a growing battle over who gets to regulate online prediction platforms in America. The federal government under the Trump administration insists the Commodity Futures Trading Commission holds sole regulatory authority, actively suing nine states to block local enforcement. Meanwhile, state officials argue that prediction sites are effectively unapproved gambling venues bypassing state laws. Donald Trump Jr. recently championed the industry before conservative state attorneys general, framing the platforms as already possessing robust federal oversight. With $1 billion in fresh capital, Polymarket secures the financial war chest needed to expand operations and fight protracted courtroom battles, but whether federal preemption will successfully shield prediction markets from state-level bans remains uncertain.

TechCrunch

Andreessen Horowitz/funding

Andreessen Horowitz expands fifth growth fund to $8.5 billion

Venture capital firm Andreessen Horowitz added $1.75 billion to its fifth growth fund, bringing the total pool to $8.5 billion after launching it at $6.75 billion in January. The expansion follows the firm's announcement of its $1.1 billion Machine Age Fund, which specifically targets artificial intelligence hardware startups focused on chips, memory, networking, and storage. Led by general partner David George, the growth investment team has backed more than 100 companies over seven years. The new funds target enterprise and consumer AI, defense technology, robotics, health tech, and infrastructure hardware and software. The expansion pushes a16z's capital deployment even further, building on the $15 billion in funding announced in January.

The rapid expansion of a16z's growth fund underscores how drastically AI has altered venture capital dynamics. Startups in the AI ecosystem are reaching mature growth stages faster than previous technology cycles, requiring massive infusions of capital at steep valuations to keep scaling infrastructure. By assembling dedicated funding for both early hardware components and massive late-stage growth, Andreessen Horowitz is positioning itself to fund the entire technology stack from silicon to consumer apps. Additionally, the firm is spending heavily on politics and lobbying during this election year to shape tech policy. It remains uncertain whether these outsized growth valuations will yield proportional returns, but a16z is placing a massive bet on persistent AI momentum.

TechCrunch

Hugging Face/lab

Hugging Face Adds Hindi and Indian English to Open ASR Leaderboard

Hugging Face and Voice Arena released Monsoon en-IN and Monsoon hi-IN, introducing the first Indic languages to the Open ASR Leaderboard. Automatic speech recognition (ASR) converts spoken audio into written text, but standard benchmarks often hide performance disparities across demographics. Monsoon addresses this by evaluating models across 4,888 speakers and tracking 12 metadata attributes per audio clip, including age, gender, geographic district, income, education, and handset brand. The Hindi benchmark incorporates a custom transcript lattice to evaluate valid spelling variants, while the Indian English dataset spans 428 native districts across 30 states and union territories. No single speaker contributes more than 6.8 percent of total duration, preventing models from overfitting to specific voices or standardized hardware.

Standard speech benchmarks usually pull convenient audio files recorded on high-end microphones by a handful of speakers, masking severe drop-offs in accuracy for diverse accents and low-cost devices. Word Error Rate (WER)—the standard metric measuring how many words a model gets wrong—can look deceptively low on average while failing completely for specific populations. Monsoon shifts how AI developers measure speech recognition by tying evaluation directly to real-world demographic diversity across the Global South. Because state lines in India closely reflect linguistic boundaries, preserving district-level location data ensures developers cannot hide systemic failures behind aggregated scores. It remains to be seen whether commercial model developers will rearchitect their training pipelines to address the granular demographic disparities Monsoon now exposes.

Hugging Face

Google Research/lab

Google Research Introduces Planetary Prediction Engine for Earth AI

Rama Pasumarthi and Shravya Shetty at Google Research introduced the planetary prediction engine, an experimental capability inside Google Earth AI that automates geospatial modeling from natural-language queries. The system splits tasks into three modular LLM-orchestrated stages: data selection, dataset curation, and AutoML training. Crucially, data passes between stages via opaque handles rather than serialized prompt text to avoid context limitations. In benchmarks, the engine hit a 76.8% mean R² across 21 CDC health indicators versus 60.0% for manual expert pipelines. It achieved 64.9% R² on FEMA risk indicators and 66.2% on the Social Vulnerability Index. Downscaling Nigerian food security models from state to local levels doubled baseline accuracy to 66.1% R². During the 2026 Bundibugyo ebolavirus outbreak in the DRC, it reached an 83.3% Recall@10, correctly flagging 15 of 18 newly invaded health zones.

Geospatial forecasting normally forces specialized data engineering teams to spend weeks curating fragmented satellite imagery, microclimate signals, and economic data. The planetary prediction engine cuts that pipeline down to mere minutes. It works primarily because it fuses explicit statistical covariates with latent representations from foundation models like AlphaEarth Foundations and Population Dynamics Embeddings. Combining clear statistical signals with complex non-linear pattern embeddings gives the system a massive advantage in data-scarce regions where coarse reporting usually hides local humanitarian crises. While these benchmark improvements over traditional Bayesian baselines demonstrate immediate value for emergency disease tracking and disaster response, the system remains experimental. It is still unclear how robustly these autonomous pipelines handle extreme out-of-distribution events, uncalibrated sensor drift, or adversarial data corruptions when deployed in real-time crisis scenarios without human oversight.

Google DeepMind/lab

Google DeepMind Releases Gemini Omni 1.1 Flash Video Generation Model

Google DeepMind released Gemini Omni 1.1 Flash, an updated video generation model available through Google AI Studio and the Gemini Enterprise Agent Platform. The model expands creative controls for video developers by introducing scene extensions, keyframe interpolation, camera control, and 4K upscaling. Developers can now extend existing video clips by evaluating up to 10 seconds of prior visual context, a substantial increase over earlier single-second context limits. Clips can be extended in 10-second increments up to a total length of 40 seconds while maintaining narrative consistency and camera movements like dolly-zooms or snap-zooms. Additionally, developers can specify start and end keyframes to generate smooth transitions and camera orbits. To lower prototyping friction, the system introduces lightweight 360p previews that render up to 60% faster at one-third the cost of standard 720p output.

Generative video models usually struggle with visual drift, jump cuts, and erratic motion when producing clips longer than a few seconds. By analyzing ten full seconds of temporal context and permitting explicit keyframe anchor points, Gemini Omni 1.1 Flash gives software engineers the predictable control required for actual post-production tools. Fast 360p draft rendering addresses another critical bottleneck by allowing creators to iterate on timing, shot composition, and camera moves without burning through compute budgets before rendering final 4K assets. While scene extensions up to 40 seconds open up realistic options for short-form media and commercial asset creation, multi-minute narrative generation remains out of reach. It also remains to be seen how well keyframe continuity holds up when interpolating between visually complex scenes with complex lighting changes, fast character action, or physics-heavy particle effects.

Google DeepMind

TechCrunch/essay

TechCrunch Opens Side Event Applications for Disrupt 2026

TechCrunch has opened applications for its Side Events program running alongside TechCrunch Disrupt 2026. The main conference takes place from October 13 to 15 at Moscone West in San Francisco, expecting over 10,000 founders, investors, and industry leaders. However, official side events can run anytime between October 10 and 16. Hosts can organize industry roundtables, venture capital office hours, deep-dive workshops, dinners, or happy hours. Approved events get listed on the official calendar and promoted directly to conference attendees. There is zero cost to apply or host, but applications must be submitted by Friday, September 4, at midnight PT with proposed dates, venue details, and target audience logistics locked in.

Organizing an official side event gives companies and venture capital firms direct access to thousands of incoming tech professionals without buying a massive conference booth. TechCrunch explicitly recommends scheduling these gatherings during the evenings of October 13 to 15, when official daytime programming wraps up and post-conference networking peaks. While TechCrunch promises to review and likely approve anything that aligns with community value, hosts still have to secure their own venues and manage operational logistics independently. The real gamble comes down to attendance. Even with placement on an official schedule, event organizers face tough competition for attendees during peak Bay Area tech week hours.

TechCrunch

Google DeepMind/lab

Google DeepMind Pilots Double-Blind AI Benchmarks to Stop Cheating

Google DeepMind partnered with OpenMined, AVERI, MLCommons, and the Singapore AI Safety Institute to run the first double-blind evaluation of a proprietary AI model. The team evaluated Gemini Flash Lite using Google Cloud's Confidential Space, a secure hardware-enforced environment inside their Confidential Computing platform. Historically, external evaluations meant choosing between two bad options: evaluators handed over secret test prompts to the model owner, or the owner handed over proprietary model weights. This cryptographic setup isolates both sides during testing. The external evaluators cannot access the model weights, and Google cannot view the test prompts. The system cryptographically proves the model was tested without human oversight or data leaks.

This mechanism targets benchmark contamination, which happens when an AI model accidentally trains on test questions ahead of time and inflates its performance score. Cryptographic isolation allows governments, cybersecurity auditors, and third-party safety institutes to stress-test high-stakes models without exposing sensitive prompts or compromising intellectual property. If adopted widely across the industry, it creates a standardized, verifiable way to audit proprietary systems before deployment. However, cryptographic security only protects the testing environment itself. It cannot fix flawed benchmark design, nor does it guarantee that third-party evaluators have crafted prompts capable of catching critical model failures or hidden vulnerabilities.

Google DeepMind

Instagram/startup

Instagram Limits Reach of Undisclosed AI-Generated Profiles

Instagram is renaming its "AI creator" tag to "AI-generated profile" and penalizing accounts that fail to disclose computer-generated personas. Under the updated policy, profiles featuring synthetic people must carry the label or face reduced organic reach across the platform. Accounts using AI merely to edit photos, polish captions, or design graphics are exempt from mandatory labeling. Meta introduced the change following user pushback and investigations into deceptive accounts. Earlier this year, a Wired probe revealed over two dozen fake male influencers promoting the dating app Goose through direct messages. Additionally, a July report by The New York Times discovered hundreds of AI-generated doctors and wellness personalities making health claims and selling supplements.

By penalizing undisclosed accounts with reach restrictions, Instagram is trying to curb deceptive synthetic influencers without banning them outright. The move comes as social networks grapple with digital identity verification, especially when synthetic personalities build real monetary pipelines off unsuspecting users. This announcement follows Meta pulling an image tool that let users generate likenesses of real people without explicit consent, as well as an eighteen billion dollar legal settlement with United States state attorneys over youth platform safety. It remains to be seen how effectively Instagram can automatically detect undisclosed AI humans, as generative images become increasingly indistinguishable from real photographs.

TechCrunch

Magna International/funding

Magna invests $35 million in Indian battery-swapping startup Yuma

Canadian automotive supplier Magna International invested an additional $35 million in Yuma Energy, increasing its majority stake in the Bengaluru-based battery-swapping venture. Yuma spun out of mobility startup Yulu in early 2023 and operates a network for electric two- and three-wheelers across 18 Indian cities, including Bengaluru, Mumbai, and Delhi. The startup manufactures its battery packs in Chennai and its charging hardware in Bengaluru, maintaining over 400 stations with more than 2,500 charging units. To date, Yuma has completed over 60 million battery swaps and deployed approximately 100,000 batteries. The company generated about 1 billion rupees ($10.5 million) in revenue for the fiscal year ending March 2026. The new capital will fund expanding swap infrastructure and doubling its battery fleet within 18 months.

While battery swapping struggles elsewhere, India's fast-growing delivery economy makes the model viable due to high daily mileage among gig workers. Fast charging takes 20 to 30 minutes and pulls drivers off the road, whereas a swap takes under two minutes. Only 10 to 15 percent of Indian gig vehicles are currently electric, leaving enormous room for expansion. Yuma remains unprofitable, but older stations are already EBITDA-positive, and management targets overall EBITDA break-even within two quarters. The business is capital-intensive because operators must build stations ahead of demand to ensure uptime. Yuma is also reducing its reliance on Yulu, expanding partnerships with five additional fleets and over 10 vehicle platforms to capture non-Yulu swap volume, which reached 20 percent last quarter.

TechCrunch

Apple/lab

Apple files new evidence against former employee in OpenAI trade secret lawsuit

Apple filed new redacted evidence in its trade secret lawsuit against OpenAI after former employee Chang Liu handed over his old work MacBook. Apple alleges Liu used a confidential Apple circuit schematic and an internal engineering application in his work at OpenAI. The filing also claims Liu enlisted OpenAI colleague Yu-Ting Peng in June to destroy evidence after learning of Apple's investigation. OpenAI defended Liu, claiming he only accessed files post-employment to help former colleagues due to Apple failing to manage residual access. Apple counters that Liu exploited a rare authentication bug. Apple is seeking a preliminary injunction to halt OpenAI's Apple-based hardware development and expedited discovery, noting that over 400 former Apple employees now work at OpenAI.

This escalation turns standard silicon valley talent poaching into a critical legal battle over hardware intellectual property. An injunction could halt OpenAI's hardware ambitions in their tracks if the court finds that stolen circuit designs laid the foundation for their devices. For hardware building, where schematics take years of specialized engineering to design, losing trade secrets directly to a competitor threatens core competitive advantages. However, the key facts remain shielded behind court redactions, including the exact nature of the circuit schematic and the internal tool name. The court must now weigh whether Apple's lost data stems from bad offboarding protocols or intentional systemic exploitation of an authentication bug across hundreds of shared engineers.

TechCrunch

Leading the Future/funding

Build American AI Plans Millions in Battleground Ads for Data Centers

Build American AI, a new group affiliated with the super PAC Leading the Future (LTF), is launching a multi-million-dollar ad campaign across Kansas, Ohio, and Wisconsin to convince voters that local data centers are good for their communities. LTF was launched with over $50 million from Andreessen Horowitz founders Marc Andreessen and Ben Horowitz alongside OpenAI President Greg Brockman. Andreessen Horowitz has directed more than $115 million into federal contributions this cycle, prioritizing AI advocacy alongside crypto PACs like Fairshake. The ad effort targets battleground states where data center construction has become a core issue in upcoming midterms. OpenAI previously attempted to distance itself from LTF in June, explicitly condemning tactics like astroturfing—where paid corporate campaigns mimic genuine grassroots support.

Data centers require massive energy and physical land, leading to significant local pushback over noise, power grid strain, and utility bills. By pouring millions directly into local advertising, venture capitalists are bypassing traditional policy lobbying to shape public sentiment at the ballot box. This strategy marks an aggressive turn in how tech firms protect their infrastructure investments. However, whether polished ad spots can overcome real neighborhood frustration over resource consumption remains unproven. It also raises awkward political friction within Silicon Valley itself. While groups like Anthropic-backed Public First Action push for AI guardrails, firms like Andreessen Horowitz are spending aggressively to eliminate friction for raw infrastructure expansion, turning local zoning fights into high-stakes political battlegrounds.

TechCrunch

Federal Trade Commission/essay

FTC Sues Amazon Over Secret Ad Auction Surcharges

The Federal Trade Commission and 22 states filed a lawsuit accusing Amazon of running a secret ad pricing scheme that inflated costs for over one million sellers across seven years. Amazon told more than 500,000 small and medium businesses that its ad auctions operated as second-price auctions, where the winning bidder pays only one cent more than the second-highest bid. However, starting in 2019, Amazon introduced a hidden surcharge called a soft reserve price. According to internal documents, Amazon used an invented auction participant—effectively a fake bidder—to drive up prices. This shill bidding tactic caused Sponsored Products advertisers to pay their full maximum bid nearly 80 percent of the time, converting a second-price model into a first-price auction.

Second-price auctions encourage advertisers to bid their true maximum value because they trust the marketplace won't extract the ceiling price unless genuine competition forces it. By secretly manufacturing fake competing bids, Amazon allegedly siphoned tens of billions of dollars in advertising revenue from independent brands who relied on that trust to budget effectively. Amazon claims its dynamic auction algorithms evaluate complex factors across billions of placements and that its pricing is standard. This case tests whether major digital marketplaces can quietly alter the fundamental mechanics of their ad exchanges without disclosure. If the FTC wins, Amazon could face massive financial restitution demands and mandatory overhauls to how transparent platform ad auctions must be.

TechCrunch

US Department of Defense/lab

Pentagon Launches Custom ChatGPT and Grok Models on GenAI.mil Portal

The Department of Defense expanded its secure GenAI.mil portal by launching ChatGPT Mil and Starshield AI's Grok for Government to 3 million military and civilian personnel. Developed under OpenAI for Government and SpaceX's satellite network initiative, these custom models are designed to meet warfighter needs without exposing sensitive data to public training sets. GenAI.mil has already onboarded over 1.7 million unique users. ChatGPT Mil focuses on unclassified routine administration, project planning, logistics, and document management through custom GPTs. Meanwhile, Grok for Government addresses broader operational tasks, targeting market research for acquisitions and real-time supply-chain management for logisticians. Notably, Anthropic's Claude is missing from the portal after the company refused unrestricted military usage without safety guardrails, leading the administration to designate Anthropic as a supply-chain risk, a label Anthropic is actively contesting in court.

Plugging commercial frontier models directly into military infrastructure marks a massive operational shift for federal defense procurement. By exempting GenAI.mil from standard consumer data collection, the Pentagon gives personnel access to fast administrative and logistical assistance while mitigating severe data leakage risks. The explicit exclusion of Anthropic highlights a growing fracture between military requirements and frontier AI labs that insist on strict safety guardrails. Providing 1.7 million defense workers with custom conversational tools speeds up daily acquisitions and knowledge management. However, questions remain regarding how these systems handle operational hallucinations during high-stakes planning. Furthermore, reliance on commercial partners like AWS, Microsoft, Nvidia, SpaceX, and OpenAI creates long-term vendor lock-in, leaving defense workflows heavily dependent on corporate technical stacks and shifting legal frameworks.

TechCrunch

Apple/lab

Phil Schiller steps down as App Store lead amid executive departures

Phil Schiller is moving on from his role leading the Apple App Store, a position he held since 2015. While he remains at Apple as a Fellow, a title reserved for significant leadership or technical contributors, his departure signals a broader shift in personnel at the company. Following CEO Tim Cook’s exit, John Ternus is stepping in to oversee a changing guard that includes the retirements of COO Jeff Williams, general counsel Kate Adams, and Apple Pay head Jennifer Bailey. Furthermore, Vision Pro lead Paul Meade has left to join OpenAI, leading to a reduction of about 100 staff members on the headset team. Carson Oliver, a 14-year veteran of the App Store, is now taking over the store’s daily operations.

This reshuffle is significant because the App Store is moving from the marketing department into the services division under Eddy Cue, changing how the platform is governed. Schiller’s tenure was defined by the high-stakes 2020 antitrust lawsuit from Epic Games, which challenged the App Store’s fee structure. While Apple successfully defended against monopoly claims, the litigation led to court orders requiring the company to allow external payment links. With so many long-serving executives retiring simultaneously, the primary uncertainty remains how new leaders like Ternus will manage the company's legal battles and product strategy. The transition of the App Store into the services division suggests a push toward greater integration of software revenue streams, but the long-term impact on developer relations remains to be seen.

TechCrunch

Hacker News

678 points522 comments

Google Has Removed MV2 Extensions from the Chrome Web Store, Including UBO

Google has officially taken the step of removing Manifest V2 (MV2) browser extensions from the Chrome Web Store, a move that directly affects legacy ad-blocking tools like the original uBlock Origin. Manifest V2 represents the older extension architecture for Chromium-based browsers, which gave developers deep access to intercept, modify, and block web requests directly in real time. By enforcing the deprecation of MV2, Google is pushing users and developers entirely onto Manifest V3 (MV3). This platform overhaul changes how extensions interact with network traffic, shifting rule execution to the browser engine itself rather than letting background scripts process requests freely.

What this change leaves open is the long-term effectiveness of ad blocking and privacy protection within the Chrome ecosystem. While Google frames MV3 around improved security, performance, and user privacy, critics point out that the rigid declarative rules API severely restricts capabilities such as CNAME decloaking and dynamic list updates. The brief announcement leaves out technical nuances about how declarative filtering handles edge cases compared to full webRequest blocking. Users seeking powerful ad blocking must now rely on stripped-down MV3 variants like uBlock Origin Lite, migrate to alternative browsers such as Firefox that retain support for MV2 features, or adopt network-level blocking solutions to protect their devices.

The discussion frames ad blocking as a critical security control against malware and scams rather than just a cosmetic preference. Commenters delve into technical trade-offs—such as CNAME decloaking limits in Chrome versus Firefox, DNS filtering via Pi-hole or AdGuard, and User-Agent spoofing—while questioning whether MV3 replacements like uBlock Origin Lite offer real protection as ad tech evolves.

articlefull thread

From the thread

beloch

Ad blocking has become a safety issue. My parents are at an age where they fall for things. If a malicious ad pops up and offers to install McAfee, some other form of crapware, or outright scamware, they'll click that ad and install the thing. Then I'll get called over when the computer starts acting screwy. If Google, MS, etc. had ever gotten together and developed a way to filter malicious ads out of their own ad services, maybe we wouldn't be having this conversation. They didn't do that and probably never will. If my parents get a new machine, the first thing I do is uninstall or, at least, delete the icons for Edge, Chrome, etc.. On goes Firefox, uBlock Origin, and a couple other extensions. They don't care what browser they use. They can't even name it. I just try to make sure they use a browser that isn't horribly unsafe. We should all do this for people we care about. Don't ask them what browser they prefer. Odds are they don't know or care. Don't try to convince them to install Firefox themselves. They won't. Just do it for them. Think of it like picking up a rusty nail that's on their lawn. Sooner or later they're going to step on that thing if you don't.

cwnythreply

The better thing would be to hold Google directly accountable for every malicious ad they allow through their network. If they cannot do that, they should not be in business selling ads then. Yes, that means amending the law (in the US). But the law as it is written facilitates crime.

siliconc0wreply

Agreed It's not just malicious software ads, they also sell ads to entities mitming every government service.

bborudreply

How would one hold Google accountable? By fining them $1bn? $10bn? $100bn? To them it is just the cost of doing business. We're beyond holding huge companies accountable. The only way they could be held accountable is if there was personal risk involved for the CEOs. There isn't.

409 points453 comments

Apple caught off guard by AI demand for Mac Mini and Mac Studio

Apple’s unexpected early product launch of new Mac mini and Mac Studio models was spurred by massive enterprise demand for local AI hardware, according to a report from The Information. Traditionally reserving desktop refreshes for autumn, Apple accelerated its schedule right before its annual iPhone event to cater to business and developer customers. The company explicitly marketed multi-system linking for Mac Studio units to run large frontier artificial intelligence models. Additionally, Apple spotlighted these desktop machines during a summer executive event with companies like Anthropic, Disney, and Ford, where the Mac mini emerged as a central favorite for enterprise infrastructure.

Despite the commercial surge, the report reveals Apple was entirely unprepared internally, lacking dedicated business engineering teams, developer relations staff, or a coherent enterprise AI strategy. When corporate buyers requested access to Apple's Private Cloud Compute infrastructure, Apple turned them down and instead relied on third-party partners like Mount Thor and WebAI. Combined with a global memory shortage, high-end Mac desktop configurations have suffered months of backorders. Consequently, frustrated enterprise buyers are opting for rival hardware like Nvidia's compact DGX Spark desktop. The article highlights Apple's struggle to manage an unplanned hardware boom while leaving open whether their long-term enterprise software support can catch up.

Commenters debate whether this surge stems from organic enterprise demand for unified memory hardware or aggressive viral marketing linked to local agent trends like OpenClaw. Tech-focused participants point out that while Macs offer cost-effective memory capacity for running local models, 32GB remains inadequate for heavy workloads, and Apple's software stack still lags behind Nvidia's CUDA ecosystem in raw compute speed.

630 points210 comments

Playa Phone

A physical phone booth at 3:30 and Ceiba near the Temple of the Flying Spaghetti Monster in Black Rock City, Nevada, lets Burning Man attendees make free five-minute international calls for a week. The creator purchased a payphone off Craigslist, gutted the coin-handling hardware, and replaced the internals with a touch-tone board and a Voice over IP adapter. Power and lighting run via Power over Ethernet. Anyone can dial +1 (775) 557-4848 from outside the desert to reach a random passerby walking past the booth, provided the line is free and someone answers before six rings.

Because cell coverage at Burning Man is historically spotty, the installation relies on an internet connection to route calls over IP. Outbound calls show up from an unknown number, which frequently triggers automated spam blockers or sends calls straight to voicemail. The project page recommends saving the number as a contact to make incoming calls work smoothly. While the hardware side is straightforward, the project leaves open exact backhaul details for the current year, which previously relied on a 20-foot tower link. The site tracks basic hourly stats and final call summaries while avoiding collecting caller identity datasets or recording call audio.

Discussions range from skepticism about wealthy burners to practical technical breakdowns and heartwarming stories. Creator aaron42net clarifies that the payphone runs over VoIP via Power over Ethernet, only tracking basic call volume metrics rather than harvesting contact data. Other commenters debate whether ephemeral, spontaneous interactions like playa weddings lose their magic if they lack legal stakes.

article

386 points196 comments

Fastpotify

Fastpotify is an open-source, lightweight Spotify desktop client built for Linux, macOS, and Windows using Rust, egui, and librespot. Instead of bundling a heavy Chromium browser engine like Electron apps do, it compiles to a native binary that launches in well under a second with minimal memory footprint. The app supports local gapless playback at up to 320 kbps and integrates Spotify Connect to control playback across remote speakers, phones, and televisions. Users can search the catalogue, manage saved playlists, browse liked songs and podcasts, and manage playback via global media keys or system tray controls.

The application includes a nostalgia-driven feature where pressing Ctrl+M transforms the interface into a pixel-accurate Winamp 2 mini player. It loads classic .wsz skin files directly from the Winamp Skin Museum, complete with a working spectrum analyser, equalizer, playlist window, and integer pixel scaling modes. While the project provides fully documented connection and credential handling under an MIT license, it leaves long-term viability open given third-party API restrictions. Furthermore, relying on immediate-mode rendering through egui means standard desktop web features like native text selection and traditional HTML accessibility hooks are absent from the custom triangle-rendered interface.

Commenters jump on the trade-offs of using egui, noting that while immediate-mode Rust graphics make UI fast, rendering everything as textured triangles breaks standard text selection and screen-reader accessibility. Others warn that Spotify is actively restricting the underlying librespot library, prompting several users to share self-hosted alternatives like Navidrome, Lidarr, and Explo for music discovery and ownership.

article

342 points195 comments

I think the military commissary's freezers were hacked

The Defense Commissary Agency, or DeCA, operates around 235 tax-free grocery stores on military bases worldwide. In late August 2026, news broke that freezers across multiple bases suffered simultaneous failures. At Fort Huachuca in Arizona, base officials confirmed on Facebook that their freezers did not merely lose power. Instead, the systems entered a forced defrost cycle, actively heating the stored food and spoiling everything inside. Official reports soon confirmed near-simultaneous refrigeration issues across at least six bases in 11 states, with earlier tallies counting up to 14 affected sites. A March 2026 DeCA procurement contract revealed that 182 locations share a central Remote Monitoring Control System (RMCS) for facility management.

While the author stops short of declaring a verified cyberattack, the technical architecture makes a centralized breach plausible. In industrial automation, Programmable Logic Controllers (PLCs) handle local equipment tasks like cycling compressors or heaters, but central network controllers like DeCA's RMCS issue top-level commands across facilities. If an attacker gains network access or an unpatched vulnerability hits the central management software, remote commands can easily force multiple units into destructive modes at once. The timing coincided with publicly disclosed industrial control vulnerabilities. Even if the incident turns out to be a flawed software patch or misconfiguration rather than state-sponsored sabotage, it highlights how network-connected infrastructure exposes critical food supplies to single-point operational failures.

The discussion brings in practical domain experience with industrial control hardware, noting that Programmable Logic Controllers and Building Automation Systems are notoriously insecure and frequently configured with default credentials or unencrypted channels. Commenters argue that a botched software update or internal misconfiguration remains far more likely than a foreign cyberattack, though some point out that low-intensity sabotage aimed at military morale would look identical to routine operational incompetence.

130 points177 comments

The safest job from AI may be writing

While programmers face cognitive burnout and potential displacement as AI agents generate code, writing remains surprisingly resilient against automation. Large language models frequently produce generic, robotic prose filled with repetitive vocabulary and predictable cadences. This failure stems from the nature of writing as a wicked problem—a systems theory term for challenges lacking clear rules, definitive formulations, or objective criteria for success. Coding involves interacting with a deterministic compiler that provides immediate feedback, whereas prose requires navigating Theory of Mind. A human writer continuously simulates the reader's mental state, managing cognitive load and adjusting arguments based on shared emotional context and lived experience, elements that statistical next-token predictors inherently lack.

The essay applies David Ricardo's economic principle of comparative advantage to human cognition. Even if AI holds an absolute advantage in raw generation speed and zero marginal cost, human effort naturally shifts toward areas with high opportunity costs where unique human traits are essential. In evolutionary biology, a costly signal represents an honest indicator of fitness because it is expensive to produce and difficult to fake, such as a peacock's tail. Authentic human writing functions as a costly signal amidst a sea of synthetic content. However, the author leaves open whether future post-training methods could bridge this gap, or if economic pressure will simply erode the commercial viability of entry-level writing roles before peak mastery is achieved.

Commenters strongly challenge the premise by pointing out that AI does not need to beat elite prose writers to decimate the writing profession; it only needs to automate the routine entry-level jobs like copywriting, translation, and technical documentation that sustain writers while they hone their craft. Others note that AI functions best as an adversarial reviewer or copilot to structured thought rather than an unassisted author, while some argue that AI models overuse cliches simply because lab training prioritizes profitable coding tasks over fine-tuned prose style.

Research

Psychology & Neuroscience

Adaptive self-organized criticality in deep neural networks

Simon Vock, Christian Meisel

Deep neural networks need to stay in a balanced dynamic state to learn well without exploding or collapsing. Researchers showed that you do not need a global supervisor to manage this balance. Instead, by letting individual synapses adjust their strength based solely on the activity of the single neuron they connect to, the entire deep network automatically tunes itself toward an optimal boundary called criticality. It works whether the starting network state is completely dormant or hyperactive, keeping signals moving smoothly across layers.

You do not need complex global monitoring to keep AI networks stable because simple local adjustments per neuron can automatically keep the whole system in the sweet spot.

Leveraging a Foundation Model for the EEG-Based Diagnosis of Alzheimer's Disease

Maggie Lin, Chung-Lin Hou, Tzyy-Ping Jung

Diagnosing Alzheimer's disease with standard brainwave scans is tricky because individual brains vary so much and standard linear tests miss complex patterns. Researchers tackled this by using a large brain foundation model, trained on over 2,500 hours of brainwave data, combined with a machine learning classifier. By feeding it just eight-second snippets of brain activity, the system reliably identified patients with dementia, reaching an impressive 89 percent area under the curve score. It outperformed traditional methods that look strictly at specific frequency bands.

An AI foundation model trained on thousands of hours of brainwaves can detect Alzheimer's signs from just eight seconds of EEG data.

Hysteresis and multistability in network spreading with neuronal activity feedback

Christoffer G. Alexandersen, Dani S. Bassett

In diseases like Alzheimer's, toxic proteins spread across the brain's structural network, while neural activity simultaneously affects and gets altered by those proteins. Previous models assumed this mutual influence was simple and proportional. The authors modeled what happens when this feedback loop is non-linear. They discovered that non-linear feedback creates distinct tipping points and multiple stable states. This means disease progression isn't always smooth; small shifts can trigger sudden jumps in disease severity, or lock the brain into long-term diseased states.

The interaction between brain activity and misfolded proteins creates sudden tipping points rather than a steady linear decline.

From the thread

nullbio

I'm convinced that this is just guerilla marketing from Apple. When this started spreading a day or two ago, it was all from no-name spam media sites that are paid to publish articles. They all claimed "a source" is where they got the intel, without specifying the source. It was spreading like wildfire on socials. The same thing happened with Mac Mini's and OpenClaw. Nobody cared about or was using Mac Mini's for OpenClaw, but there were all of these very suspicious posts from accounts that were clearly Apple marketing bots (you could tell by looking at their post history, where they would drop "Mac Mini" into every conversation they could across all different subreddits and unrelated topics). Then it became fairly common. So Apple's marketing department is seemingly using the same strategy again. Because still, it's impossible to find a reputable source for this claim.

srslackreply

I am using Mac studios for local inference, but not for openclaw. It was simply the best bang for the buck at the time, and the software side of things has gotten even better since then. I haven’t run the pricing comparison but I have a feeling with the direction they’re heading that it’s better to hold off for m6 as it may assuage various bottlenecks if you are latency sensitive.

bredrenreply

Reportedly, the M6 will not get pro, max or ultra editions. Instead Apple will go straight to M7 for these.

pylotlightreply

Some reporting seems to reject that btw: https://wccftech.com/apple-m6-pro-m6-max-soic-mh-wmcm-packag...

ipsum2reply

Apple's marketing isn't that smart. They're very traditional. You give them way too much credit.

mgh2reply

Additional source: https://www.youtube.com/watch?v=OpqGf0m5FaQ

api

Apple has a huge opportunity here to lead the market for machines to run local models if they step into it. Their stuff is already better than what nVidia is offering with stuff like the DGX Spark. It's a niche market but it's a market that overlaps heavily with professionals in the AI space and lead developers, so it's a market that gets them customers in those roles. If I were running Apple I'd call the RAM price bubble for what it is and temporarily eat some margin to offer machines with more RAM than competitors, especially these models that are great for edge AI, and capture market share.

bigyabaireply

Apple doesn't design GPUs on-par with Nvidia's efficiency yet. They need an architectural overhaul to be a serious competitor, which is what I'm expecting is queued up for M7. Nvidia has CUDA, AMD has CDNA, and Apple has... compute shaders, I guess?

niccereply

> Apple doesn't design GPUs on-par with Nvidia's efficiency yet How much it matters in inference? Most GPUs have enough computing for that and the bottleneck is the RAM speed and size. And M5 Ultra is becoming to challenge this.

bigyabaireply

For prefill, it's basically all that matters. Long-horizon agent tasks, session compaction, file reads and context manipulation will all hit the compute bottleneck in regular usage, incurring several minutes of latency on most Apple Silicon chips, regardless of RAM. It's kinda why memory bandwidth is an enormous red herring, even for datacenter applications. Nvidia's huge advantage is a compute-optimized GPU architecture and their Infiniband networking, their memory controllers aren't really the star of the show.

Danoxreply

Probably two generations away. I’m more interested in how much uplift/speed and more importantly what is the power usage is required for the new computers Apple is shipping particularly for the Studio versions.

Danoxreply

Apple isn’t the company that eats margins but they are company that would design around the problem and I think that’s what they will do after all, they have the design and engineering and plenty of money because they didn’t burn it on AI models or data centers.

mwcampbellreply

> Their stuff is already better than what nVidia is offering with stuff like the DGX Spark. How so? In tokens per second when running major open-weights models, or something else?

alasdair_

There is a lot of "AI demand" that isn't just running inference on an LLM whose weights you downloaded. I'm training a model using reinforcement learning with self-play. I can and do use vast.ai when scaling but for experiments it's far faster, and cheaper, to run it locally until the bugs are all figured out. Just provisioning a new instance and copying the relevant checkpoints and things can take 25 minutes. It's zero locally.

kridsdale1reply

Likewise. I have a huge demand personally to run AI noise-filtering models on many TB per month of raw video files. It takes about 3 days per file. Apples ProRes codec is only licensed to run in high quality mode on a Mac, and so my Nvidia PC can’t do what I need. Thus, I own the beefiest Mac Studio you can currently buy. I would pay more for more TFlops. I have done local LLM on there but it wasn’t interesting. Far worse performance and intelligence per dollar than the cloud boys. There is no cloud offering for my video needs though.

varispeedreply

Out of curiosity, how would you transfer many TB to cloud and back if such service was available?

firecallreply

My guess is that it wouldn’t be TB sized files. It’s probably files that, over the course of a month, add up to multiple TBs. Which would suggest that a 1Gbit fibre connection would be adequate. For serious commercial usage, multi Gbit fibre is available in many places around the world. Although being video files, they could easily be in the TB range. In which case, it would be interesting to know!

tassadarforaiurreply

Have you looked into those tinycorp nvidia drivers for thunderbolt egpus on Mac?

ashish01reply

What are you training on using self play? Like alpha go? Curious what your setup is like .

alasdair_reply

Yes, basically like alphago. I’m teaching it to play magic: the gathering. I had to start with some heuristic-based bots that played the decks very simply just to get to the point where the was some signal to learn from. I did behavioral cloning on the bots as a foundation, then self-play.

jmalickireply

Are you training something so big you need that much unified RAM though? If you can fit it on a GPU, and especially for training, it is so much quicker than a Mac.

Xeoncross

If I had to pick a product, I'd say an affordable 32GB mac would be the sweet spot for running local models that function well like Qwen 3.8. It's true, most people don't run models, but being the default platform for running open weights seems like it has plenty of advantages right now. Just like sales benefited from developers defaulting to MacOS for most open source languages like Ruby, Go, Rust, and TypeScript.

tristorreply

> If I had to pick a product, I'd say an affordable 32GB mac would be the sweet spot for running local models that function well like Qwen 3.8. 32GB is not enough RAM. I don't even own a device with less than 36GB at this point, and that device I only have because my employer is being cheap. 64GB is a reasonable starting point for running local LLMs + normal tasks. 128GB let's you really run most smaller models like Qwen 27B and 35BA3B with good context. Even Qwen3.8-Flash-Next runs in 128GB with a 4-bit quant. 32GB would be limited to running models like Gemma4 12B and smaller dense Qwen versions like 9B unless you were using very small quants which damages quality of response.

Xeoncrossreply

You are mistaken. I'm running Qwen 3.7 28B 4bit (MLX) with a 200k context window and everything total is 32GB RSS. Is this the best? No. That's why I said the sweet spot. Getting from 16GB macs to 32GB is perhaps possible. Jumping to 64GB or 128GB as the default is simply unreasonable right now.

e28etareply

I have a similar machine, and briefly poked at running a local LLM, but got discouraged after a couple days. The quality, responsiveness, and impact on the rest of the system didn’t seem worth it to me. What sorts of things are you doing with the local LLM? Anything interactive? Should I take another look?

tristorreply

I assume you mean Qwen 3.8-27B? Yes, you can run this in 32GB of RAM, but it's very context limited. With KV cache compression and other techniques, it's better now than in the past, but I'd still want more RAM, personally. EDIT to add that you need to reserve 8GB for the system if you don't want to cause problems on macOS, which means 32GB RAM = 24GB max for model + context. It takes 18-19GB to load a 4-bit quant of Qwen3.8-27B, so I'd be really surprised if you can actually get a 200k context window. You need to fit within a 24GB WSS (which is generally a more constrained RSS) to get stable performance on 32GB RAM.

Foobar8568reply

Memory used : 38GB, and I haven't even started a LLM nor podman, I always fight with memory when using LLM on my mac with 48gb. And I don't remember to have been able to have pushed to 200k context Qwen 3.6. 3.8 is running on my RTX 5090.

mirekrusinreply

32GB is not enough, it's unified/shared memory, you need to have space for usual system and user apps/services. 64GB+ or dedicated 48GB (2x24 on GPUs) is IMHO absolute minimum.

redox99reply

32GB of fast unified memory is enough for Qwen 3.8 27B. - 16GB for the weights at Q4 - 9GB for the full 256K context at Q8 - 7GB spare for overhead and system. The problem is that these Macs have 32GB of slow unified memory. Edit: I'm thinking of a headless Mac mini, if you meant running it on the same machine you're using of course you'll need more memory, but LLMs are best served from a headless server so that's what I'd recommend.

hawk_reply

Is this for setup for agentic coding? Why not also run the IDE compiler etc... on the same machine to use those CPU cores as well?

qeternityreply

> Edit: I'm thinking of a headless Mac mini, if you meant running it on the same machine you're using of course you'll need more memory, but LLMs are best served from a headless server so that's what I'd recommend. What? LLMs are best served from a massive PD disaggregated cluster of B300s connected via NVLink. If you're running LLMs on a Mac Mini, it's because you want to run local, not because it's the best setup.

shagiereply

> If I had to pick a product, I'd say an affordable 32GB mac would be the sweet spot for running local models that function well like Qwen 3.8. https://www.canirun.ai (five months ago: https://news.ycombinator.com/item?id=47363754 377 comments)

From the thread

solfox

Our first time out on the playa together, my girlfriend and I stopped by this phone booth and made a couple of calls. And because we stopped, we saw the FSM camp next door, and a giant sign saying "Weddings Here", so we inquired. They offered us free rings and then announced our wedding over a loudspeaker. In 2 minutes we had an entire wedding party, complete with best man, an elder to give away the bride, 25+ guests, and an officiant. My newly-minted best man even gave me one of my favorite playa gifts ever as a wedding gift. We opted not to "make it official", though they offered. It was a truly incredible experience. Thanks FSM!

k7peakreply

That's awesome, congrats!!! Amazing things happen on the playa.

tucnakreply

What's the point if not to make it official? Otherwise there are no stakes, no benefit, and the whole thing would amount to cosplay, right?

freeplayreply

Oh I thought make it official meant consummating the marriage in front of everyone. If it meant making it an official marriage, I agree.

virgil_disgr4cereply

what's the point of going to burning man? What's the point of going outside? What's the point of reading? what's the point of posting a comment on HN?

nilamoreply

What's the point of going to a murder mystery diner party, if no one gets arrested at the end?

sebastiennightreply

I love how your idea of the failed murder dinner party seems to still include an actual murder ; the problem is only the absence of an arrest.

charcircuit

I bet the dataset of who is calling who is pretty valuable. It's people related to others who were willing to drop a bunch of money to go to the desert, so you know they likely have some disposable income. If this operator isn't collecting this, it may be a good business opportunity for a competitor.

jspaetzelreply

Slow your roll, not everything is a startup, some things are just for fun

charcircuitreply

The best kind of things are both fun and profitable enough to make them financial viable to keep doing.

aaron42netreply

The only data I collect is my phone bill, and the only thing I do with it show hourly stats on the site and a summary of calls at the end of the event.

perarneng

0 outgoing, nobody knows numbers anymore and even if you could look it up you won't punch numbers like its 1992. With that said, amazing project. super cool that people can call in

aaron42netreply

It has changed a bit in recent years, but there's still only very spotty cell service at Burning Man. Enough people can figure out how to look up a phone number that this phone still dials out thousands of times during the week. The bigger problem for this project is because of phone spam, fewer people will answer calls from unknown numbers, so a lot of those calls go to voice mail.

retiredreply

Is it possible that people don’t realize you can make free calls? Perhaps they think you need coins, which they don’t have on them. Similar to interactive installations in museums. People are afraid to touch things in museums. You really need a big sign to tell them it is the purpose to play with it.

ape4

I looked but I don't see how it works. How does it get power? I suppose it uses a some single board computer and VOIP.

MarkMarinereply

there are a lot of gas generators powering people's camps.

aaron42netreply

I ripped the coin-aware guts out of a pay phone I bought from Craigslist, and replaced them with a basic touch-tone board, then hooked that up to a voice-over-IP adapter so it can make calls over the Internet. All of that is powered via power-over-ethernet, including the light in the sign above the phone. It used to get a little bit of Internet directly from Burning Man itself via a 20 foot tower pointed near center camp, but now it just uses Starlink, which is easier for my campmates to set up without me and much more reliable. For power, it is now plugged into a solar generator that is occasionally recharged from a gas generator.

From the thread

alexyoung

I realise I’m pissing into the wind here, but I find the LLM text on the homepage and docs quite funny/ awkward. > “Ctrl+M turns it into a tiny player that wears any classic Winamp 2 skin, spectrum analyser, equalizer, and playlist included. 2000s vibes, pixel for pixel.” Everything is said with too much intensity, and phrasing that sounds impressive but doesn’t really mean that much. Like “wears any classic Winamp 2 skin” feels so awkward, what’s “wears”? You mean it can use it? I feel like if you can’t be bothered to write the code, at least document it yourself and write the marketing copy so I know you understand the product. Otherwise how can I trust running it on my computer? Did the LLM generate some amazing rm -rf somewhere, or another blunder that wrecks data I might care about?

doublerabbitreply

I too enjoy wearing the RealPlayer skin of the 90's. Makes me look very chic.

bryanrasmussenreply

If the last twenty years of horror movies have taught me anything it is that maybe you can't be a real player, but you can always wear their skin!

jcarranoreply

If it can transform into Winamp I don't care if the code is LLM generated, typed by infinite monkeys or found in some ancient scrolls. Jokes aside, how can you be sure that the official Spotify app (or any app, or Windows) was not written by a LLM without supervision?

phatfishreply

The official Spotify app has some skin in the game. Random dudes self promotion project, not so much.

hk__2reply

Not OP but I don’t really care if it was or not. All I know is that I’ve been using the Spotify app 10+ years and it works well. I don’t know if this random app works, and if it hasn’t been supervised I doubt it does.

MrGilbertreply

Interestingly enough, https://fastpotify.rocks/winamp/ shows what I dislike about LLM-generated text: From "The buttons do what they say [...]" to "[...] a different format and are not supported" is just a huge wall of text without any visual guidance. It's as if the person writing these lines didn't care about the result. One could argue that people not caring about written documentation isn't anything new, but in the past, these documentations also had little text. So they were easier to digest. Now, we get great walls of text that are hard to read. I cannot say that I like that.

alasanoreply

Not running code from someone who has so little self awareness that they publish documentation like this tbh. Doesn't matter if it's open source.

abrookewood

I think the key call out with this is how cool egui is (the Rust GUI library it uses). Check out their website: https://www.egui.rs/ The demos are very good: https://www.egui.rs/#clock Edit: egui is an immediate mode GUI library written in Rust. egui runs natively on , and on the web it is compiled to WebAssembly and rendered with WebGL or WebGPU. Everything you see is rendered as textured triangles. There is no DOM, HTML, JS or CSS. Just Rust.

WithinReasonreply

The anti-aliasing is broken, it doesn't apply gamma correction

IshKebabreply

Did you look at it recently? Until a few months ago text was a blurry mess but it seems like they've mostly fixed it now.

trouporeply

> Everything you see is rendered as textured triangles. There is no DOM, HTML, JS or CSS. Just Rust. So that's why text selection doesn't work.

Fumblenutsreply

This feels like it would be an accessibility nightmare? But maybe it could be cool for doing some advanced animations or rendering games or something.

wilted-iris

Spotify is in the process of killing the librespot project that this and most third party Spotify players are built on. I think the golden age of music streaming is coming to an end. I’ve migrated to a self hosted library with streaming and radio for discovery. I hope we’ll see many projects in the space flourish. Many, like Navidrome and the whole OpenSubsonic ecosystem, seem to be doing quite well.

nan60reply

> radio for discovery Would you be willing to share a bit more about this setup? The only thing keeping my on streaming is the fact that its almost impossible to find new music otherwise. I know LastFM exists but it just doesn't click for me the way that streaming does. I like discovering music organically.

dylan604reply

> radio for discovery was a big question mark I had as well. The Clear Channel and iHeart stations are not known for discovery. Now if you find a college station or by some miracle an independent station you can get some new stuff. My favorite local college station is also ad free. They also have an app which I lean on as it's not the strongest transmitter so their coverage is spotty. Maybe there's other indy stations with apps as well??

BLKNSLVRreply

This radio station is local to the city I live in, and they've been great for music discovery for me, although heavy focus on local artists, but they've put me onto a few international artists as well. https://threedradio.com/ I subscribe to them, have done so for ~10 years I reckon. I don't listen as much as I used to, but I want them to continue existing either way. Be aware, the quality of their playlists is incredibly variable, but that's part of the charm. I guarantee you'll hear something you haven't heard before... for better or worse. Example: https://www.youtube.com/watch?v=6zHR0HMFY60 (with 2 minutes of silence at the start to, presumably, avoid auto-copyright claims).

mcmcmcreply

Public radio is great, NPR and the member stations have some great streaming channels. KEXP and KCRW are two favs

wilted-irisreply

I believe there are! I consistently hear good things about KEXP and FIP but I haven’t listened to either extensively. I listen to a lot of SomaFM (free, https://somafm.com/) and the Flow State radio-podcast (paid, https://www.flowstate.fm/).

bubblemothreply

I left spotify around two years ago for a self hosted option (I mostly use navidrome+feishin). Music discovery was a concern for me as well, but then I realized that I didn't like the idea of spotify curating which music I discovered. So far, this is what has worked for me: - Actually listening to the music I have now. I think the art of listening through full albums has slowly died, but there is so much music we're ignoring because we just listen to Artist's top tracks. - Word of mouth. Asking my friends what they listen to. This is a way that people used to discover music and I've found so much good music this way. - Bandcamp. Bandcamp is SUCH a good way to discover new music. You can listen to the music in app, there are great articles and you can actually support the artists!

pixelatedindexreply

Epic bought bandcamp many moons ago. Did they enshittify it or is it still going strong?

fbnlsrreply

I have pretty much the same setup. The only thing that bothers me with navidrome+feishin is that gapless playback is not working perfectly. Also, a great resource for discovering new music is the public library. They often have a curated CD section that I then rip into FLAC. I've had great chats with the people in my town's library, and I often pick a totally random CD off the shelves and discover some amazing artists.

yard2010reply

While moving from spotify to youtube music a few years ago I bought a vinyl player to listen to full albums and fell for it. Then when I moved to curate and self host my collection with Plex and Plexamp I realized I'm not using the player anymore since I'm using Plexamp to listen to albums. Plexamp is not perfect and not open source unfortunately, but it's a really good non-enshitified product. I had to migrate away from youtube music since the experience was infuriating me every time. I love making playlists on the go. In youtube music after adding 30 tracks to the play queue, tapping 1 track accidentally deletes the play queue and plays this track. This made me feel bad every time it happened. I don't see how it is just accepted to lose progress in 2026 as if we're using win95. It feels like 80% of the sw I use provides a caveman UX, not leveraging the tech, leaking implementation details and being constrained by technical reasons. On the other hand, working on Prepbook[0], I realized how hard it is to create the perfect experience and how it's even harder to justify it financially. I'm doing discovery by talking to people, I love it. Every time I listen to Monolink I recall my co worker that just said "check out monolink" when he heard I love Ott. I love that it makes the experience about connection to people as well. [0] https://prepbook.app

msmaghreply

using nts.live and and fip radio -> https://www.radiofrance.fr/fip at the moment for same purpose. Radio France allows to save liked track names in profile to look up later.

lolivereply

In addition to those two bangers, BBC6 and RNE also have shows to discover new musics. PS: RadioParadise is also a safe place for good music.

pbasistareply

> Spotify is in the process of killing the librespot project How could they possibly do that? They cannot prevent people or AI from reverse engineering their APIs.

TingPingreply

Depends on how much they care. They can take down repos containing any keys, or use heuristics to block clients.

hahn-kevreply

Technically no, but they can just ban you. It would not be impossible to detect.

egeozcanreply

Anthropic can prevent people from using their models with the subscription plan outside Claude Code, why couldn't Spotify?

sbinderreply

Because millions of Spotify-compatible speakers are sold with firmware that will never be upgraded, so they can't ever change the Spotify connect protocol without breaking those devices.

protocolturereply

>I think the golden age of music streaming is coming to an end. Already ended for me. They killed the plan I was on, and then started sending me begging emails to sign up for a new, more expensive one. But once I experienced their platform with ads, I just stopped using it entirely.

Walf

Fuck Spotify. Even if you don't care about the ethics of Daniel Ek's investments, the platform itself is crap. One of the lowest paying to artists, not paying at all unless they hit a threshold, and the catalogue is filling up with AI slop, which they push on users via 'radio' and 'mix' playlists, so they can pay even less royalties. No, thanks.

SuperMousereply

I've tested Deezer for half a year. The software quality was horrible. Had a lot of crashes, it often took many many seconds to change between the songs and often getting stuck, the buffering algorithm was nuts and eating a lot of bandwidth. Everything on a standard Pixel. Full 5G coverage was available. Alternatives?

Walfreply

Tried Tidal? The one gripe I have with it is no remote control, as in using phone to change what's currently playing on the device connected to my sound system, without it thinking I want to start playing on my phone. Apple Music pays artists better, too.

cmehdyreply

Tidal is the last streaming service I use, split between that and a jellyfin+arrs setup coupled with buying music from artists directly. Fuck Spotify and their rotten leadership. Was amongst the early adopters (2009). I'm sure they won't care for the loss, and I won't miss th when they're gone.

joriswreply

Seriously tried Tidal. Gave up on it because of the software quality.

zcrar70reply

Qobuz have the highest per stream payout rate for artists [0], and have the best audio quality of all streaming platforms (better even that Tidal imho). They have desktop, mobile and android auto (and presumably carplay) clients, and an equivalent to Spotify Connect whereby streaming can be switched to a different client. Downsides are no native client (same webview approach as Spotify), not as strong recommendations (though slowly improving), good but slightly quirkier UI (also slowly improving). For me the trade offs are worthwhile. [0] eg. https://resources.onestowatch.com/ethical-music-streaming-pl...

ronjakoireply

I've been happy with Qobuz as well. I use the third party open source QBZ client on Linux desktop.

mcmcmcreply

I left when they forced podcasts into my music app and never looked back

tamimioreply

The best is to host your own, navidrome is great and easy to set up. But if you really wanna hate on one, it’s Apple Music, believe it or not, they will actually delete ALL your playlists that you made over the years if you stop paying subscription for few weeks or sometime for days, why? Because Apple says fuck you, you are renting and you are not entitled for anything, which also explains why they are now trying to have hardware as a service with their new renting hardware program.. at least spotify doesn’t do that and keep your playlist to export later.

From the thread

gwbas1c

To summarize for people who TLDR: 14 freezers failed at the same time. They are all internet-controlled, and failed at the same time as a disclosure about a vulnerability. They all failed by turning on the defrost cycle and heating food. Regardless if this was a hack or a bug, the bigger lesson is that overcomplicated systems fail in catastrophic ways. Why do military commissaries need remote-controlled freezers? It seems like a very fragile, and needless, way to run a freezer. --- But, there are some options that the author didn't consider: 1: This could be a quickly applied patch that failed. 2: This could be a "script kiddie" hack from someone who isn't a government actor. I'm less onboard with a state actor. Generally, when a state actor has hacked something, they don't want the victim to know. In this case, if it was a state actor, I would anticipate that they would make a single freezer fail in a way that they could verify using something like a hacked video camera or otherwise by watching public social media feeds. IMO: A state actor would only "make sense" if they knew the hole was closing soon and they don't care if they're discovered, perhaps because their operation is winding down.

quickthrowmanreply

These aren’t your typical refrigerator or freezer, these facilities have walk-ins or purpose built cold storage with multiple evaporators and condensing units. A building automation system is pretty standard for most buildings above a certain size, and monitoring and controlling the refrigeration is usually a part of it. Unfortunately, I would wager that all BAS software is full of flaws and holes, allowing access to it for the public internet seems like a bad idea. I need to be on my company VPN to access our locally hosted BAS front end (which I have authorized access for) which seems like the bare minimum security.

stackghostreply

>Generally, when a state actor has hacked something, they don't want the victim to know Could be the Iranians, or someone aligned, conducting anti-morale operations. Could be the start of a series of small but annoying failures.

TeMPOraLreply

Or could be a rouge LLM in one of the big labs, that accidentally self-prompt-injected itself with the title of that vulnerability research paper.

stackghostreply

I suppose. But if I was an Iranian operative with instructions to damage the morale of the American war machine, I'd absolutely do a series of things like this. It doesn't really harm anyone, but enough similar events and the families of servicemembers start to feel unsafe, which is psychologically very different than the servicemembers themselves feeling unsafe. It's probably more benign than that, most likely firmware or a central controller failure, but still

elictronicreply

Iran is actively looking for ways to attack back against the United States especially against military targets without actually escalating the situation. I'm sure some script kiddie broke into a government network, hacked an industrial process, and forced a limited supply piece of equipment into a failure mode that takes some thought and is more unique as an attack vector. It's just like buying hacks for CS source right?

peterabbitcook

A couple years ago I worked on a service that had to communicate with a Siemens S7-1500 PLC. Based on my experience with that project, none of what I’ve read recently about unsecured industrial PLCs is surprising. I opened Siemens TIA Portal and PLCSIM for the first time and thought “wow, I didn’t think the Windows 95 GUI library was still supported.” None of the PLC contractors we had hired knew how to enable TLS on the thing (user/pass eg admin/admin was their usual). Anecdote: I once spent hours reading the docs and clicking around trying to get it to accept an SSL certificate signed by a real CA and it wouldn’t go, but it accepted one I self-signed in openssl. In all fairness, the people who are experts in the field of Siemens PLC programming are usually mechanical-ish engineers and security is not in their skill set or on their mind.

sidewndr46reply

Isn't this the industry expectation in that kind of equipment? If it was signed by a real CA the cert. could expire and render the equipment unable to communicate.

stephbookreply

also you can't pin the user/pw to the machine with a note, because someone might need remote access. better stick with admin/admin

peterabbitcookreply

That’s a tough question. If your PLC is on an airgapped LAN, admin/admin is not great security hygiene but you’ve reduced most of the risk by airgapping. On my project the service I wrote was doing bidirectional communication with the PLC over OPCUA. The server running this pod was connected to the internet, so it was critical to have proper TLS for the OPCUA client/server. Rotating LetsEncrypt certs on the system every 45 days is a lot of toil, but using a self-signed cert that expires in 2040 from some dev laptop doesn’t pass muster in most organizations either. That’s just the OPCUA path.. In these projects You also typically see WinCC HMIs that can talk to the PLC from anywhere on the network without TLS. And also SIMATIC Web Server pages - minimally secured by default, and the amount of info you can grok about the system in Chrome Dev Tools is troubling.

dylan604reply

> Rotating LetsEncrypt certs on the system every 45 days is a lot of toil What is unique about your system using LetsEncrypt that you can't automate certbot to handle this task as it was designed and intended to be done?

sidewndr46reply

I'm not sure if you're speaking from personal experience, but most I've interacted with don't have to worry about the self-signed vs. LetsEncrypt debate. They just don't do it. Also there would be no way to do LetsEncrypt as the system is air gapped.

gopher_spacereply

> admin/admin is not great security hygiene but you’ve reduced most of the risk by airgapping. admin/admin is on a post-it stuck to the monitor because everyone we hire is perfectly capable of rooting a machine they have physical access to.

tgsovlerkhgselreply

My guess would be something like the CA using some feature that was newer than what the equipment would support (e.g. ECC signatures but the equipment only supporting RSA), not an intentional "no real CAs" decision.

katzenversteherreply

Most factories I know do not allow their PLC be accessed from the internet. They are usually on a separate Network. However, the "engineering" station (the computer running e.g. TIA Portal) sometimes is. The PLC engineers I had contact with usually had an electrical engineering background. That's why they like PLCs in the first place with the ladder logic programming languages, grafcet and if they feel fancy a bit of structured text (assembly like) or structured control language (pascal like). They indeed did not know much about software security but a great deal about machine safety. A real security nightmare are older OPC servers (OPC-DA) which is super reliant on DCOM. OPC is quite important to connect the PLCs to SCADA systems or 3rd party devices.

lowbloodsugarreply

>They are usually on a separate Network. Then someone plugs in a cable because boss wants something "over there" and there's already a network that runs "over there". Or optimizes to a smart switch with vlans, and then someone else optimizes to a single vlan. It's not hard to not give a shit, or not understand, network security.

lenerdenatorreply

That someone can be brought into an office and shown a small diagram of the approved network topology. Then they can be shown a small diagram of the current network topology. Next, they can be asked if they're the same. If they're not, they can finally be asked if they're aware that deviating from the approved network topology without consulting infosec is grounds for termination of their employment.

jordanbreply

My mind was blown when I realized that the way tftp works is that as the machine is booting it asks the network if anyone has some software for it to run.

Joker_vDreply

Well, what else can it do, really? It has to boot with pretty much zero knowledge about the external world (maybe except asking the user for the current date and time). Sure, you can hardcode an outdated list of CAs (it's always outdated because the system can be booted 10/20/100 years after it was made) in but that just opens you to unexpected obsolescence, and you usually can't put too much stuff in the bootloader anyway. Not really dissimilar from the human upbringing: leave a baby with "bad" guardians, and it will grow up corrupted. That's a feature, not a bug: if you knew what behaviour exactly you wanted (other than "whatever Simon says"), you would just bake it in in the first place, right?

Dylan16807reply

If we pretend we're revising TFTP boot in 1995, let's have it get up to 20 boot options from the server and their md5 hashes, and if it's not set to auto it waits for the user to pick one. It then verifies the hash as it downloads. Also it uses TCP for the download.

CobaltFire

As someone who spent over 20 years active duty, and spent a ton of my career in the IT, security, etc. side of the house: Unlikely to be a hack, more likely to be a misconfiguration or update sent incorrectly. That said, the timing of the disclosure and the issue are rather concerning. Regarding the highest value targets to hit with an attack like this, you would want to target Guam, Hawai'i, and other isolated overseas locations where this would have ripple effects in the local economy. Guam specifically would cause catastrophic supply shortages, since DeCA probably supplies around 50% of the groceries on that island (that's a WAG based on my time there).

ericmayreply

Generally agree with your assessment, but in the case of Guam or other more remote installations if there were catastrophic issues we'd just airlift food in. Costly but certainly manageable. Hawaii I'm not sure why that would be an issue unless the whole island was attacked or shut down. Even if the on-base shops were hacked you could just go shop at Wal-Mart or Costco or any number of other locations on the islands. If there was an extended issue then the commander could authorize meal stipends as they do for some units today and then you would just go buy food off-base. Ideal? No. Manageable? Very much.

scheme271reply

More to the point for Guam and similar locations, canned and non-perishable food can are probably around and more can be airlifted in to get people through it. It was pretty much the standard back in the day.

avs733reply

The diet in Guam is already heavily dependent on shelf stable foods. I know this from personal experience but theres a surprising (to me) amount of research on it as well (c.v., https://www.guampedia.com/health-consequences-of-modern-diet...). However, it is worth noting that food on a military installation =/= food nearby, for many obvious and non obsvious reasons.

cyanydeezreply

With the current admin, assumptions about capabilities are just wrong. One of the reasons for consitency above all else is to use the power of history. Ahistorical is now the defacto standard. All asdumptions about America power are toilet paper.

nwatsonreply

Did they airlift better supplies to a floating aircraft carrier somewhere in the northern Indian Ocean recently? No.

nkriscreply

Surely airlifting supplies to a carrier in a war zone is not totally comparable to airlifting supplies to a base in a US territory?

nwallinreply

Aircraft carriers are not particularly well suited to be supplied by cargo aircraft. The largest cargo aircraft with the strengthened landing gear, arresting hook, and catapult attachment used to be the C-2 Greyhound. It had a range of approximately 1,500 miles with a 10,000 pound cargo. It was retired on July 28th, 2026. The C-2 has been replaced by the CMV-22B Osprey. It has a range of only 1,150 miles and a more limited cargo capacity of only 6,000 pounds. 6,000 pounds isn't nothing, but it's not a lot either. There's not a lot of stuff within 1,150mi of the Arabian Sea. Your two options are one of the US bases in the southern Persian Gulf (Kuwait is too far) and Diego Garcia. Neither are ideal. In both cases you're doing a transfer out in the field. On the other hand, the C-5 Galaxy can carry 120,000 pounds to a range of 4,800 miles. You can fill it with food in the continental US, fly to Hickam AFB in Hawaii, refuel, and fly the rest of the way to Guam. On the return leg, an unladen C-5 can fly directly to the 22nd Airlift Squadron's home base California without a layover. The C-17B Globemaster is smaller than the C-5, although it's nearly as capable. With a 160,000 pound cargo, it has a range of 3,200 miles, (the C-17A has a shorter range) which is just short of the 3,800 miles to get from Hawaii to Guam. I don't know how much cargo it's capable of carrying 3,800 miles, but on the order of 100,000 pounds is a reasonable guess. It is also capable of an empty flight back to the US without a layover.

psunavy03reply

You're just revealing your ignorance of Navy supply chains with this comment, and I say that as a veteran who has no support for the ongoing shitshow, only sympathy for my former colleagues trapped underway.

TZubirireply

>Unlikely to be a hack, more likely to be a misconfiguration or update sent incorrectly. Are these materially different? Assuming that there wasn't a literal command to turn off all freezers, from an organizational, non criminal perspective, to the organization the damage will be the same, the root cause will be a bug (whether exploited by chance or malice), and the fix will be the same(fixing the bug).

gavinsyanceyreply

> Are these materially different? A hack implies an adversary, and intentional targeting.

TZubirireply

But is it materially different? Or is it a difference in quantitative increase of likelihood of the exploit? The thesis "everything hackable" will be hacked, and with the advent of hacking LLMs, the boundary between criminal hacking and civil hacking is being blurred, it's no longer attributable to a criminal intent, hacking commands are just something that happens through no criminal intent. Practically, the hackers, whether human or syntehtic, are a catalyst, they accelerate the rate at which a bug that causes damage will manifest, and perhaps the magnitude of the manifestation, but the essence of the negative manifestation of that bug is the same. Whether it was triggered by a solar ray flipbit or by a rogue LLM or by a russian hacker, the essence of the issue is the same.

segmondyreply

I always wonder by folks with some "inside" knowledge like you will then come out to share more details. Why? I understand that there's no security through obscurity, but I don't think that the details to get your point across matters. Loose lips sink ships.

munificentreply

We are a very social species who likely evolved to feel good when we share juicy information.

CobaltFirereply

Nothing I shared is, in any way, even close to something I am concerned with sharing. This type of wargaming scenario is literally stuff they give to brand new kids out of high school to get them thinking. It's not anything close to "loose lips sink ships."

0xWTF

This is eerily suggestive of a vulnerability Hank Paulson hinted at in his 2014 book "Dealing with China" "Every nonelectric cooler comes with 25 years of free real-time monitoring. On a visit to the company in the spring of 2012, I watched as technicians in Broad Air’s space-age control room checked on the performance of its units in locations as diverse as the Adolfo Suárez Madrid–Barajas Airport in Spain; Qualcomm headquarters in San Diego, California; and Fort Stewart, the U.S. Army base in Georgia [emphasis added]. "Zhang says that 80 percent of his clients are repeaters. “If you bring long-term benefits for your clients, they will choose you.”"

Terr_reply

America's tech-sector has a similar problem, which--until recently--was tempered by the idea that it was a dependable and predictable ally to most of its customers. Not just in the sense of secretive kill-switches, but "US government commands you to turn over this encryption key and you're not allowed to say you did so" stuff.

vkoureply

> "US government commands you to turn over this encryption key and you're not allowed to say you did so" That barely even registers given the scale of more modern escalations against its 'allies'. Escalations like "US government threatens that you aren't a real country and that you can't defend yourself from it."

From the thread

pshirshov

No. Below is my emotional opinion based on my own experience: Currently the models are totally helpless with plot and emotions. They make epistemic, logistical and temporal mistakes. But: They are very good in finding _your_ epistemic, logistical and temporal mistakes. They are great as adversarial reviewers. They can validate grounding. They can test voices. They can build complex parallel reconstructions representing timestamped inner monologues, dialogues and events. A process consisting of many models representing characters + states + epistemic isolation + reviewers could be great to validate or disprove ideas. Models can write you formal models of your writing. They can run solvers against your formal models to - again - validate and disprove ideas. Literally, you could validate your writing with TLA+, SAT solvers, queries against formal ontologies. They can build corkboards of unimaginable complexity. A moderately good writer who plays well as a part of human-machine writing apparatus could be a competitor even for a top-tier author. When you write with a model, you define disciplines. You write your prose. You let your model turn your prose into formal systems and verify it. You let your model criticize you. Sometimes you let your model to polish your crude language into something sleek. So, at current state of LLM development, writer as source of ideas and governing process is safe. Writer as a vision-expansion machine is not. I've been very hostile to LLMs in literature. I've started to write a short novel about that. As a joke I tried to use models after half of the text was done. I've applied my engineering skills. That changed both me and my opinion. Models are awesome if you know their advantages and weaknesses. A so-so writer with a good model and a good approach to prose development could produce epic stuff.

fpolingreply

If it will be true, then we should have many superb books by new authors at this point. But we do not. This is similar to observations that effect of AI on quality of apps in stores is mostly non-existent.

svachalekreply

Neither book stores nor app stores pay for quality, beyond some minimum low bar. Markets are optimized for getting product out quickly, sales and marketing, and follow on effects of succeeding with a hit. If you made a better app, game, book, or song you need to get through the noise of millions of competitors and for the most part, it's never going to happen. In short, the expected effect of AI will be more stuff, faster, not better. Are we seeing that? I think so.

nottorpreply

> A so-so writer with a good model and a good approach to prose development could produce epic stuff. Well, there have been hugos/nebulas awarded to "creative workshop" quality work before (and i mean before LLMs). But that doesn't make those books "epic".

pshirshovreply

You can write an epic piece and never reach any sort of broad audience. Not mutually exclusive.

tptacek

I appreciate what this piece is trying to do directionally but it shares a logical flaw with lots of other pieces about AI and knowledge work. It assumes that in order to disrupt a particular field, AI needs to print serviceable work unassisted; that it's "vibe-shipping". AI today is most effective when it's not vibing, but rather copiloting a skilled operator. I don't necessarily want my agent to build an entire system for me, even if it's ultimately the author of almost every line of code; I'm actively making decisions throughout. There's obviously a spectrum here but at most points on the spectrum the amount of assistance available is still a step change in the economics. So too with writing. The first rule of accelerating writing with AI is that you're not allowed to use a single word the AI suggests. Even if what the model comes up with is great, better than what you could have done, as soon as the AI suggests it it's poisoned. At least with current models, readers can detect LLM prose in the parts per trillion, and as soon as they do you've lost them. The second rule of writing with AI is that AI encouragement is toxic. A structural consequence of RL is that models are exquisitely tuned to generate responses that make their users perceive value. We recognize this in a gross sense in "sycophancy", but the problem recurs fractally in at finer-grained levels, where stuff like "this part is really strong" will subtly allow the model to set a course for your writing and you'll confidently ship crap. With those two rules in mind, models are incredibly valuable for writing, more valuable in my experience than the professional copywriters I've worked with. The trick is to get them to make suggestions at a higher level than just writing alternatives: * Do the sentences in these paragraphs end with the new idea or information? * Are the real actors in each sentence the grammatical subjects? * From paragraph to paragraph is there a clear flow of topics, or are things jumping around? * Is this piece crudded up with metadiscourse like "it's important to note"? I've had a stack of notecards for ages that I took down from Joseph Williams "Style: Towards Clarity

matherial

I strongly disagree, for reasons that are different from what's discussed in the thread. It doesn't matter if LLMs can't hold a candle to some of the finest human writing. What matters is that before you become a world-renowned writer, you need to pay your bills, often for a long time. And that's what LLMs take away. All the mundane literature-adjacent works: journalism, translations, technical writing / corporate comms, copyediting. The same goes for many other forms of art. A budding painter or a musician could support themselves off commissioned / commercial work while working on their grand opus... but now, the customers just prompt gen AI.

hparadizreply

I find myself writing more than ever. In fact I must. I have to explain in no uncertain terms what I mean. Though I often find myself in very complex technical straits where words are not enough. There's something liberating about writing a wall of text and knowing that at least something will give you feedback. It won't be annoyed. It won't be tired. It won't say "tldr?" back at you. I don't know if that will convert into a novel for me. But it must for someone.

johnnyanmacreply

Yeah, the arts in general in this age are reverting back to 1) the privilege of the rich and secure who can do it despite the lack of financials. 2) the ultra passionate who spend every waking moment outside of their day job engaging in the craft. Often to the detriment to their health and social obligations. You shouldn't need to go all in on a skill just for a chance to maybe one day earn money from it.

deepwoodsreply

I generally agree, but don't discount the impact of platforms like Patreon and Substack. Obviously, many authors of the past had patrons, but this was generally based on status and connections - and that's to say nothing of people who can support a writing career based off inherited wealth. But there is little historical precedent for an avenue for creatives to support themselves (or at least meaningfully supplement their income, potentially leading to less non-creative work hours and thus more time to hone the craft) using a decentralized base of support.

enraged_camel

The reason LLMs use the same cadence and cliches is that their RLHF does not emphasize writing well, the way it emphasizes, say, coding. And the reason for that should be pretty obvious: coding is where the big money is at. So the author may be correct, but for a different reason: unless writing starts being very valuable as a profession, it's unlikely the labs will spend significant resources making their AI models better at it.

jimmyl02reply

What's interesting is that we still read AI's output and that's still the primary interface between a human and the AI. Implicitly that means that writing is important and I think we're already starting to see this with the complaints around how Claude "talks"

ElProlactinreply

Claude "talk" is particularly annoying if you're using Claude to perform a job function (write code, analyze a spreadsheet, etc.) but average people reading an ad, Facebook post, etc. don't really care from what I can tell. In fact, I think most of them have no idea what Claude "talk" is, even if some of them notice common patterns that have emerged recently.

swatcoderreply

No, they use "the same cadence and cliches" because they inflate a short and ambiguous prompt into long and specific prose by making statistical assumptions about what best fills in the gaps. It's not a training problem, it's an information theory problem, and it's not really surmountable. Any given model will always have some distinct implicit voice that its biased towards for that infill content, and so a popular model will always become exhaustingly common, painfully familiar, and cliche. Users can use more elaborate prompts that shift the voice away from the most normative and towards some other nodes, but they need to put in special effort for that, and what people-at-scale specifically want from these tools is to put in very little effort, so we can expect that overwhelming number of casual and naive users will always be generating cliche slop with them. Code escapes this problem not because of training but because it specifically benefits from cliche (boilerplate, patterns, etc) and so an model whose code "voice" reflects your own taste as a coder (or your toolchain's taste as a vibecoder) is going to feel like productive output rather than slop. But it's still cliche.

famouswafflesreply

No. Reinforcement Learning is doing a lot here. Anyone who played with these models before the Davinci intstruct-tuning (completion) era can tell you the same. In some ways, SOTA models have gotten better at writing, but the neuroticism of instruct-tuning has still not been resolved.

KineticLensmanreply

"The reason LLMs use the same cadence and cliches is that their RLHF does not emphasize writing well, the way it emphasizes, say, coding. And the reason for that should be pretty obvious: coding is where the big money is at." Software can be checked for being 'written well' by compilers / linters etc. There is no equivalent for well-written natural prose. Spelling and grammar checkers haven't a clue about prose semantics.

ilakshreply

They can use human writers and also LLM-based multi-step scoring systems with analytical guidelines.

v2026.09.01

Weather

World Clocks

Markets

v2026.09.01

Weather

World Clocks

Markets

The Debate

If the British Parliament had accepted the Conciliation Plan of 1775 and retained the American colonies, would chattel slavery in North America have ended sooner than 1865?

Does constitutional judicial review derive its legitimacy from popular sovereignty, or does it inherently conflict with democratic legitimacy?

Is human social order fundamentally maintained through internalized shared norms or through institutional coercion?

From the Labs

Google Research releases TimesFM-3 for zero-shot multivariate forecasting

Alteon raises $2.5 million to build long-endurance wind-powered aircraft

Polymarket raises $300 million from Donald Trump Jr.'s 1789 Capital

Andreessen Horowitz expands fifth growth fund to $8.5 billion

Hugging Face Adds Hindi and Indian English to Open ASR Leaderboard

Google Research Introduces Planetary Prediction Engine for Earth AI

Google DeepMind Releases Gemini Omni 1.1 Flash Video Generation Model

TechCrunch Opens Side Event Applications for Disrupt 2026

Google DeepMind Pilots Double-Blind AI Benchmarks to Stop Cheating

Instagram Limits Reach of Undisclosed AI-Generated Profiles

Magna invests $35 million in Indian battery-swapping startup Yuma

Apple files new evidence against former employee in OpenAI trade secret lawsuit

Build American AI Plans Millions in Battleground Ads for Data Centers

FTC Sues Amazon Over Secret Ad Auction Surcharges

Pentagon Launches Custom ChatGPT and Grok Models on GenAI.mil Portal

Phil Schiller steps down as App Store lead amid executive departures

Hacker News

Google Has Removed MV2 Extensions from the Chrome Web Store, Including UBO

Apple caught off guard by AI demand for Mac Mini and Mac Studio

Playa Phone

Fastpotify

I think the military commissary's freezers were hacked

The safest job from AI may be writing

The Rabbit Hole

Lake Ontario

Research

Adaptive self-organized criticality in deep neural networks

Leveraging a Foundation Model for the EEG-Based Diagnosis of Alzheimer's Disease

Hysteresis and multistability in network spreading with neuronal activity feedback

From the Bookshelf

The Innovator's Dilemma

Disruptive Innovation

Contrarian Corner

Fine-tuning an AI auditor to detect hidden backdoors can teach it to ignore those exact internal signals.

Spaniards are rapidly abandoning their famous late-night dining culture in favor of earlier dinner hours.

Feeding private corporate communications into AI models can unintentionally make automated cyberattacks easier.

Photograph

On This Day

Games

Puzzles

The Three Boxes

The Escalator Step Count

The Shifted Cipher

The Track

Cariño

Quote

Word

Komorebi

Albedo