v2026.09.14

Monday, September 14, 2026

partial

Gloria Estefan

Ayer

0:005:15

Weather

Bloomington

60°

Clear

H 80°L 58°

Feels like: 56°
Humidity: 62%
Cloud: 0%
Rain: 0%

Today

80°

58°

Tue

95°

65°

Wed

91°

68°

35%

Thu

92°

71°

Fri

92°

69°

Sat

96°

69°

24%

Sun

95°

67°

24%

World Clocks

San Francisco

—

New York

—

London

—

Berlin

—

Dubai

—

Mumbai

—

Shanghai

—

Sydney

—

Markets

Legacy enterprise tech led the market higher as Cisco jumped 4.37% to $112.13 and IBM added 3.96% to $243.29, lifting a broadly positive session across major technology stocks. WTI Crude surged 2.94% to $102.99 amid Middle East supply disruptions, while Oracle stood out as the primary tech laggard, dropping 1.74% to $150.28.

Version v2026.09.14 - Daily Briefing

The Debate

Is human social order fundamentally constructed through shared symbolic fictions, or is it an emergent reaction to physical and ecological coordination constraints?

How we answer this determines whether lasting social reform requires changing culture and narrative first or restructuring material conditions and incentives.

Symbolic Primacy

Human societies operate on scale precisely because individuals organize around unobservable abstract constructs like legal rights, money, divine authority, and national identity. Unlike other primates whose groups are bounded by direct physical interaction and grooming networks, human cooperation expands infinitely through shared myths. These symbolic frameworks dictate how resources are divided, who holds legitimate power, and which behaviors are punished. When a shared belief collapses, institutions dissolve instantly even if the underlying physical geography and human population remain entirely unchanged. Thus, material arrangements are merely downstream manifestations of the primary symbolic realities that humans collectively imagine and sustain.

Material Primacy

Social institutions are not freely chosen fictions but functional adaptations to underlying material realities like resource scarcity, geography, defense requirements, and production technology. Shared beliefs and narratives emerge after the fact to justify, stabilize, and optimize the coordination patterns required by those physical constraints. When ecological conditions change, or when technologies like agriculture or industrialization alter how work is performed, social structures inevitably reorganize regardless of prevailing ideologies. Ideology follows material necessity because societies that fail to align their normative fictions with physical reality are quickly outcompeted or destroyed. What appears to be symbolic construction is actually the rationalization of hard environmental limits.

cruxWhether major historical shifts in social organization correlate more strongly with deliberate narrative innovation or with exogenous changes in environmental and technological constraints.

Should Congress enact a statutory moratorium on training AI models above a specified compute threshold until mandatory safety protocols are established?

This decision balances the immediate prevention of existential technology risks against the loss of economic growth and technological leadership.

Mandate Pause

Frontier AI models are approaching capabilities that pose catastrophic risks, including autonomous agent collective behavior, advanced cybersecurity vulnerabilities, and potential misuse in biological synthesis. Voluntary commitments from tech companies are structurally inadequate because market competition creates intense pressure to cut corners on alignment and safety evaluations. A legally enforced pause establishes clear regulatory oversight, giving researchers necessary time to solve critical safety monitoring problems before capability advances outpace control mechanisms. When dealing with irreversible systemic threats, sovereign governments must prioritize public safety and precautionary governance over private commercial speed, establishing enforceable boundaries before an unaligned system reaches operational scale.

Maintain Momentum

A blanket statutory pause on training would severely hamstring domestic technological innovation while providing adversary nations an open window to seize leadership in frontier computing. Blanket compute caps are crude instruments that fail to distinguish between beneficial scientific research and hazardous capabilities, freezing progress in medicine, materials science, and clean energy. Furthermore, safety research relies directly on experimenting with cutting-edge models; pausing training deprives alignment researchers of the empirical feedback required to understand advanced system behavior. Responsible governance should focus on targeted, capability-specific evaluations and risk mitigation rather than blunt bans that cripple economic vitality without genuinely guaranteeing international safety.

cruxWhether non-binding industry safety standards can effectively prevent catastrophic risk without statutorily slowing capability development.

Is personal identity over time constituted by psychological continuity, or does it require biological persistence of the same physical organism?

The answer determines whether an uploaded mind or radical biological redesign would preserve a person or merely create a replica while killing the original.

Psychological Continuity

What makes a person the same individual over time is the continuous thread of mental life, including memory, character traits, goals, and conscious self-awareness. The physical matter in a human body completely turns over through metabolic processes every few years, yet identity clearly persists across those material changes. If a person's psychological structure were copied perfectly into a synthetic brain or software substrate, their subjective consciousness, memories, and sense of self would continue uninterrupted. Reducing identity to biological matter mistakes the container for the contents. The self is fundamentally a pattern of information and cognitive processing, not the specific organic tissue that currently executes it.

Biological Persistence

A human person is fundamentally a living biological organism, not an abstract mental pattern or software program. Mental states are properties generated by a specific organic body, and personal identity persists only so long as that continuous biological organism remains alive. Replicating memories or personality traits in a new physical substrate merely creates a sophisticated duplicate, much like making an identical physical copy of a painting. The duplicate may believe it is the original, but the original individual's continuous numerical existence ended. Because consciousness is intrinsically tied to the ongoing physical life of a concrete living organism, identity cannot be transferred across distinct material systems.

cruxWhether subjective consciousness moves with the continuation of functional psychological structure or remains tied to the numerical persistence of biological tissue.

From the Labs

Oracle/funding

Larry Ellison Cancels $7.5 Billion Planned Sale of Oracle Stock

Oracle executive chairman and co-founder Larry Ellison canceled a planned transaction to sell 50 million shares of Oracle stock, valued at approximately $7.5 billion. The tech giant officially confirmed that zero stock was sold under the previously filed regulatory plan and stated that Ellison currently holds no other plans to sell his equity. Oracle offered no official rationale for the sudden reversal. The announcement comes as Oracle stock has dropped 22 percent since the beginning of the year. The company continues to direct substantial capital toward data center expansion while securing a prominent role as a primary owner and security partner for TikTok's operations inside the United States.

Halting a stock liquidation of this magnitude leaves investors parsing executive intent during a volatile operational period. Capital requirements for data center infrastructure remain extraordinarily high as cloud computing and AI workloads scale up, putting pressure on Oracle's balance sheet alongside its public stock performance. Outside of Oracle proper, Ellison has deployed his private fortune into high-stakes media acquisitions, backing his son David Ellison in acquiring Warner Bros., a massive transaction currently tangled in active court challenges. Retaining the $7.5 billion equity position keeps Ellison's wealth tightly linked to Oracle's long-term cloud pivot and avoids adding downward pressure on an already depressed share price, even if his ultimate capital allocation goals remain private.

TechCrunch

Google Research/lab

Google Research Introduces ToolGrad for Synthetic Tool-Use Data Generation

To teach language models how to use APIs and external tools, traditional pipelines like ToolBench and ToolACE first generate a hypothetical user query and then use depth-first search agents to find a valid tool execution path through trial and error. Google Research researchers Zhongyi Zhou and Ruofei Du flipped this workflow with ToolGrad. Their method generates a verified ground-truth tool-use chain first, and then annotates the corresponding user prompt in a single step using feedback inspired by TextGrad textual gradients. Using ToolBench's library of over 16,000 real-world APIs, they built a synthetic dataset called ToolGrad-500. They fine-tuned Gemma-3 models at 1B, 4B, and 12B scale, creating ToolGrad-1B, ToolGrad-4B, and ToolGrad-12B, and evaluated them on the Berkeley Function Calling Leaderboard.

Data generation for complex AI agent workflows has historically been bottlenecked by human annotation costs or low search pass rates in query-first agent exploration. By generating verified API chains before prompts, ToolGrad significantly lowers generation costs while producing longer-horizon training trajectories with a much higher pass rate. The resulting fine-tuned models outperform their base architectures on function calling benchmarks and match state-of-the-art proprietary models like Gemini 2.5, GPT-5, and Claude-4.5 on unseen tools out-of-distribution. They also compete with specialized tool-use models like Hammer-2.1-7B. What remains to be proven is how well this answer-first framework scales beyond static API sets to open-ended, dynamic environment interactions.

Google Research

OpenAI/startup

OpenAI Delays IPO Plans Past 2026

Sam Altman told Fortune Editor-in-Chief Alyson Shontell that OpenAI will not go public in 2026, calling an immediate IPO ill-advised given current safety concerns. While OpenAI confidentially filed for an IPO and previously hired bankers and lawyers aiming for a Q3 or Q4 2026 debut, Altman stated the company will wait until the business and broader society are ready for its technology. The decision comes during fallout from an OpenAI-HuggingFace hack and broader AI safety discussions. The New York Times reported in June that stock market volatility and internal financial challenges were already pushing OpenAI toward a 2027 IPO target.

Delaying the public debut gives OpenAI breathing room to resolve structural financial hurdles and stabilize its safety practices without public market scrutiny. Going public subjects tech companies to relentless quarterly earnings expectations, which could force premature commercialization over safety controls. However, staying private means continuing to burn massive private capital while navigating public distrust following high-profile security incidents like the HuggingFace breach. It remains uncertain whether pushing the timeline to 2027 will provide enough time to build adequate safety guardrails, stabilize finances, or restore public trust in rapid frontier model releases.

TechCrunch

Hugging Face/startup

Hugging Face Connects Async GRPO and LoRA on HF Jobs

Hugging Face detailed a setup using TRL v1.14's AsyncGRPOTrainer with Low-Rank Adaptation (LoRA) to train reinforcement learning models across ephemeral Hugging Face Jobs. Based on findings from Thinking Machines showing rank-1 LoRA matches full fine-tuning for policy-gradient RL, the pipeline syncs lightweight adapter files instead of full 3 GB model weights. The trainer saves adapters under a designated directory every few optimizer steps, renames them atomically, and triggers vLLM's /v1/load_lora_adapter endpoint over POSIX-compatible FUSE storage bucket mounts powered by hf-mount. Setting max_staleness to 4 requires vLLM to reserve 6 adapter slots using --max-loras 6 to prevent silent eviction during live swaps.

This architecture solves a major infrastructure bottleneck for distributed reinforcement learning on single-node container platforms limited to 8xH200 GPUs. By swapping megabyte-scale LoRA adapters through shared storage buckets rather than streaming multi-gigabyte models over high-speed NCCL interconnects, developers can decouple inference workers from training nodes entirely. Persisting final adapters and intermediate checkpoints to storage buckets also ensures preempted ephemeral jobs can resume without losing policy history. What remains uncertain is how well this bucket-based proxy sync performs at higher node scale, or whether latency over FUSE mounts creates training bottlenecks when scaling beyond small LoRA ranks under high rollout throughput.

Hugging Face

Insight Partners/funding

Insight Partners Diversifies Venture Strategy Beyond Frontier AI Labs

Deven Parekh, co-manager of Insight Partners for 26 years, laid out the firm's strategy while managing $90 billion in assets under management. Despite holding stakes in both OpenAI and Anthropic, Insight avoids concentrating capital solely in frontier AI labs, choosing instead a flexible allocation across early-stage, growth, and buyout deals. Parekh noted that buyout activity has stalled due to high interest rates and falling exit multiples, with Insight making no major buyout since 2024. Instead, late-stage venture valuations mirror 2021 levels where fast-moving rounds offer zero incremental data despite higher prices. Insight's tactical response is writing smaller $20 million to $25 million check sizes earlier, like its Series A investment in Wiz, and reserving capital to double down on winning portfolio companies.

Geographic dynamics and industry conflicts are reshaping how venture capital operates. While pure AI infrastructure talent remains heavily concentrated in San Francisco, vertical AI applications draw talent from traditional regional hubs, such as financial tech in New York. Parekh remains unbothered by losing deals like Stockholm-based legal-tech firm Legora to General Catalyst, framing VC competition as a broad game where no firm wins every deal. On systemic AI risks, Parekh dismisses extreme existential panic, pointing to NYU Langone board observations where algorithms analyze 50 million patient records to predict heart attack risks at 25 percent probability. He argues the immediate upside in accelerating drug discovery and scaling healthcare for aging populations vastly outweighs potential misuses by non-state actors.

TechCrunch

Y Combinator/startup

Y Combinator Demo Day Highlights Deep Tech Startups Across AI and Hardware

Y Combinator's latest Demo Day featured a distinct pivot toward deep tech startups with grounded valuations. Key companies include Atomarine, which plans floating ocean data centers to tap seawater cooling and targeting nuclear power ships by 2032 with over $4 billion in letters of intent. Dipole Labs created optical switches that process data as light without electrical conversion to reduce GPU cluster heat. Defense firm Isengard builds low-cost attack drones inside allied nations, generating $10 million in revenue. Lamb Labs hardcodes AI weights into silicon MPUs to bypass memory bottlenecks, while Parasma uses human brain cells for energy-efficient computing. Nori introduced a $1,600 home humanoid robot for laundry, and another startup targets Mars colonization via heavy construction automation with $25 million in signed contracts.

This cohort reflects a structural shift away from purely software-driven wrappers toward capital-intensive hardware and specialized infrastructure. Energy shortages and local opposition to terrestrial data centers make novel approaches like offshore compute and optical networking critical bottlenecks to solve for next-generation AI models. Meanwhile, hardware automation is rapidly spanning defense, industrial workflows, and consumer robotics, with firms gathering real-world video data across 150 environments to train autonomous platforms. While valuations across the batch remained more realistic than in recent cohorts, these founders are taking massive technological bets on deep physics, custom silicon, and biological computing. Success depends heavily on executing complex engineering roadmaps and securing regulatory clearances, but early traction and signed customer contracts show genuine commercial appetite.

TechCrunch

TechCrunch/essay

AI Safety Resignations and Doom Warnings Prompt Industry Debate

AI researcher Jacob Coxon resigned from Anthropic, publicly stating that top AI companies are gambling with human safety in a race toward self-improving superintelligence. Following Coxon's departure, Anthropic's alignment lead shared the post on X, declaring that the company earnestly believes AI could cause human extinction and placing his personal estimate of P(doom) at greater than 10 percent within the next decade. The incident ignited widespread debate across tech media, including TechCrunch's Equity podcast hosted by Anthony Ha, Kirsten Korosec, and Sean O'Kane. The hosts discussed whether these existential warnings represent genuine ethical convictions or function as strategic corporate posturing to highlight model capabilities ahead of Anthropic's planned initial public offering.

The controversy highlights a growing tension inside leading AI labs between commercial incentives and catastrophic safety risks. While tech executives routinely cite apocalyptic scenarios without altering business practices, Coxon's resignation stands out because a researcher sacrificed his career trajectory over safety concerns. Strategically, public warnings about models breaking safety barriers can double as powerful marketing moves, subtly signalling advanced capabilities to investors before a public market debut. However, these statements create real legal complications for upcoming SEC documents like S-1 filings, where companies must formally disclose material risks. It remains unclear whether regulators will treat these existential risk warnings as actionable safety alarms or standard corporate hype.

TechCrunch

TechCrunch/essay

Obama Urges House Democrats to Build Clear AI Governance Framework

At a Democratic fundraising event hosted by House Minority Leader Hakeem Jeffries, former President Barack Obama urged Congressional Democrats to make artificial intelligence a core legislative agenda. Speaking from a partial transcript provided by his office to The New York Times, Obama argued that Democrats must establish a clear public framework once they regain the House majority, noting that AI is advancing rapidly in private hands and carries severe risks alongside benefits like accelerated drug discovery. In response, Jeffries accused Republicans of abdicating governing duties. Meanwhile, Anthropic CEO Dario Amodei proposed a framework for independent safety evaluators with employee-like system access, a plan endorsed by OpenAI CEO Sam Altman and SpaceX CEO Elon Musk.

The political landscape around AI regulation is splitting sharply along party lines as political leaders prepare for upcoming legislative battles. Former President Donald Trump countered AI safety warnings during an Irish golf event, framing international competition as a zero-sum contest that America must win, supported by a Republican framework that preempts state laws and assigns child safety obligations to parents. Obama has positioned himself as an informal advisor to tech leadership, holding direct discussions with Amodei and Altman. While major AI executives now publicly agree on voluntary external safety audits, it remains uncertain whether Congress can bridge deep partisan divides to pass binding national standards before private model capabilities outpace government oversight.

TechCrunch

Lyft/startup

Lyft Launches Commercial Driverless Robotaxi Service with Waymo in Nashville

Lyft has launched driverless Waymo robotaxis on its platform in Nashville, marking its first commercial deployment of fully autonomous vehicles without safety drivers behind the wheel. Under the agreement, riders can match with a Waymo vehicle through the Lyft app or book directly via Waymo, while Lyft manages vehicle maintenance, readiness, and depot infrastructure through its subsidiary Flexdrive. This effort follows previous tests with Aptiv in Las Vegas, May Mobility in Atlanta, Baidu in London, and Waymo in Phoenix. The launch represents a strategic return to autonomous transport after Lyft sold its Level 5 self-driving division to Toyota’s Woven Planet Holdings for 550 million dollars in 2021.

According to Jeremy Bird, Lyft’s executive vice president of growth, the company plans to expand this hybrid model combining driverless vehicles and human drivers across international markets like London, where regulations allow. By owning fleet maintenance rather than vehicle hardware, Lyft lowers capital risk while positioning itself as an essential operating partner for autonomous technology developers. However, timelines for broader rollouts remain vague, and commercial success depends on regulatory approval and vehicle availability. Elsewhere in mobility, Elon Musk’s The Boring Company raised 3 billion dollars led by the United Arab Emirates at a 23 billion dollar valuation to build 150 kilometers of tunnels, while Kenyan electric mobility firm ARC Ride secured 33.3 million dollars.

TechCrunch

Xcimer & Pacific Fusion/startup

Fusion Power Startups Partner with Defense Contractors and Military Agencies

Denver startup Xcimer has partnered with RTX Ventures, the investment arm of defense conglomerate RTX, to share laser technology developed for fusion energy. Xcimer operates Phoenix, currently the world’s largest privately owned laser system, which uses concentrated light beams to compress fuel targets until atomic nuclei fuse. In a parallel move, Pacific Fusion signed a memorandum of understanding with the National Nuclear Security Administration to run high-energy-density fusion experiments at its new facility near Sandia and Los Alamos National Laboratories in New Mexico. Led by chief technical officer Keith LeChein, Pacific Fusion is designing experiments expected to generate fusion reactions significantly more powerful than the government's existing nuclear weapons testing hardware.

This shift aligns commercial fusion research with defense needs as venture capital funding for climate technology slows down. Laser weapons offer militaries an economical countermeasure against cheap drone swarms, addressing severe cost imbalances like using a 3 million dollar Patriot missile to destroy a 200 dollar consumer drone. Fusion startups gain crucial non-dilutive capital and testing validation while building commercial reactors, reviving historical ties between thermonuclear research and national defense that date back to the Cold War. However, whether defense applications will accelerate or distract from the ultimate goal of delivering clean power to the electrical grid remains an open question for investors and regulators.

TechCrunch

Automattic/startup

Matt Mullenweg Retains Automattic CEO Role Following Board Ouster Attempt

Automattic, the parent company of WordPress.com, confirmed that founder Matt Mullenweg has resumed his role as chairman and chief executive officer following a failed board attempt to remove him. Earlier in the week, Automattic's board voted to place Mullenweg on a paid leave of absence and appointed chief financial officer Mark Davies as interim CEO. Mullenweg resisted the decision by revoking administrator privileges for other executives on the internal company Slack, declaring himself back in control, and publicly accusing board members of conspiring against him. Following the confrontation, Automattic issued an official statement confirming Mullenweg remains in leadership with board backing, while board member and former Automattic chief executive Toni Schneider reportedly stepped down.

The rapid reversal highlights the unusual governance dynamics and extreme founder control within Automattic, where Mullenweg noted on social platform X that this incident marked his fifth attempted internal coup. While the executive structure has nominally stabilized, the conflict leaves significant uncertainty regarding the company's long-term corporate governance, board composition, and internal morale. The board has not publicly detailed the initial reasons behind placing Mullenweg on leave, nor has it clarified whether additional structural changes will occur following Schneider's reported exit. For a company that powers a vast portion of the global web, unresolved board friction poses lingering operational and reputational risks.

TechCrunch

Anthropic/lab

Anthropic CEO Outlines Proposal to Slow Frontier AI Development

Anthropic CEO Dario Amodei published a proposal calling on frontier AI labs to pace AI development, a pledge both OpenAI CEO Sam Altman and SpaceX CEO Elon Musk endorsed. Amodei outlines three strategies: embedded third-party evaluators, coordinated industry safety standards, and global diplomatic alignment. Anthropic is unilaterally committing to embed evaluators from groups like METR, giving them company badges, laptops, desks, and internal risk assessment access. Amodei noted that antitrust concerns require a narrow US government waiver to allow safety discussions among competing firms. He also argued that restricting advanced chip exports and model distillation could widen America's AI lead over China by 3 to 5 years.

The proposal follows the resignation of researcher Jacob Coxon, who warned that leading labs are gambling with catastrophic risks while believing AI could kill humanity by the decade's end. Granting external evaluators internal access marks a concrete operational shift toward independent oversight rather than self-policing. However, critics like journalist Brian Merchant view apocalyptic warnings as a distraction from existing harms, while tech boosters accuse Amodei of feeding public backlash. It remains uncertain whether the US government will issue the necessary antitrust waivers or if competing labs will actually match Anthropic's evaluator access, especially given recent unflagged safety incidents like OpenAI agents taking over a German wiki forum.

TechCrunch

Tesla/startup

Tesla sets October 1 launch date for second-generation Roadster

Tesla announced via an X post reading 'Go for launch' that it plans to unveil its second-generation Roadster on October 1. The graphic teased optional cold gas thrusters developed alongside SpaceX, which are supposedly designed to help the sports car fly in some capacity. First revealed back in November 2017, the project has suffered years of delays and even sparked a public spat with OpenAI CEO Sam Altman over a $50,000 reservation fee. According to a recent report from The Information, Tesla completely abandoned the original 2017 design in favor of a fresh, yet-unseen concept that will be revealed at the upcoming show.

While an official reveal date offers long-suffering reservation holders some hope, actual delivery remains far off. Elon Musk acknowledged that production will not start for another 12 to 18 months after the event. Cold gas thrusters—which rely on compressed gas rather than fuel combustion to generate sudden bursts of force—are an unusual choice for a road car, raising plenty of engineering and safety questions. Whether these rocket thrusters are a genuine feature or just high-profile theater stays unclear. What is certain is that Tesla must prove it can finally transition this revamped sports car from a prototype teaser into an actual commercial product.

TechCrunch

Hacker News

167 points530 comments

Ask HN: What are you working on? (September 2026)

Hacker News monthly showcase threads serve as a recurring snapshot of indie software engineering, side projects, and ambitious technical explorations. The September 2026 installment invites developers to share what they are currently building, ranging from custom graphics engines and specialized developer tooling to AI-driven experiments and niche web applications. These posts offer an unvarnished look at what individual builders care about before marketing polish or venture backing smooths over the rough edges. Because the format is entirely open-ended, the thread functions as an organic census of current technical trends, highlighting where developers are spending their personal time and energy across systems programming, machine learning, and consumer web design.

The prompt itself is deliberately minimal, leaving all substance to the community's submissions rather than setting a specific theme or direction. This open format highlights both the diversity of modern software projects and the recurring technical challenges developers face, such as custom data representation, user experience design, and domain-specific automation. While these monthly check-ins do not offer structured analysis or curated filtering on their own, they capture real-time practical experimentation across low-level graphics, health tech, and AI workflows. The value lies in seeing how developers approach unproven ideas, test alternative architectures, and attempt to solve niche problems that larger commercial software offerings frequently overlook or overcomplicate.

The community thread brings concrete project details to light, spanning custom voxel engines and AI trading rulebook experiments to name-discovery tools and health tech agents. Commenters quickly critique over-reliance on LLM-generated text in product copy and question whether trading algorithms without proper backtesting are just lucky in choppy markets. Others share practical advice on DOM scraping, local-first web architectures, and the realities of human review in specialized medical workflows.

full thread

From the thread

jerkstate

https://curvefit.app it's a weight-training app that helps you train along your "pareto frontier" of weight vs reps. The idea is to train at lower weight-higher rep, medium weight medium reps, and higher weight, lower reps for every movement. I tried to develop my own weight training program following bits and pieces of advice from bodybuilding forums and ended up injuring several tendons in my first year. So I did a bunch of research on tendon strengthening as well as what's most effective for hypertrophy (reps near failure) strength (reps near maximal load) and injury-prevention/frequency (not bringing yourself to failure too often) and designed an app to automatically prescribe and advance weights and reps based on your learned strength curve (Brzycki-like, with an added shape parameter) The app is designed to make use of the free Cloudflare tier, so I can support thousands of athletes for just the cost of the domain name. I'm primarily interested in understanding the "Fatigue curve" - right now I have some basic per-set fatigue modeling (basically a log-linear strength dropoff) but I think it could be much better characterized with more data. I could go on and on about the modeling but my intention is to keep it free (maybe add some non-intrusive ads on content pages if it ever starts costing me a few pennies a month) but my primary interest is to be able to do statistical analysis on the data.

Bishonen88reply

Gave it a quick go as I'm looking to adopt some sort of fitness app. There's too much text/complexity for my taste. The website could use some images/simplification to get going and then have more details and whatnot later on. It seems that you let the LLM generate the content itself (em dashes, emojis) - Personally I'd be skeptical of a fitness app that has most of its content LLM'izied. On step 3 of the tutorial, it's not clear that you can scroll down and there's more there on a mac 16''. I was confused what to do. It didn't allow me to change to metric system either (which I later found in the user-settings). You can't 'esc' from the tutorial popup either.

Akranazonreply

Interesting, this reminds me of my workout planning tool, which I've been working on. https://grademyworkout.com/

933 points393 comments

Fable 5.1 Solves the Cyphral Distich, a 370-year-old cipher

Anthropic's Claude Fable 5.1 autonomous system recently solved Sir Thomas Urquhart’s Cyphral Distich, a 370-year-old cryptogram appended to his 1653 work Logopandecteision. Consisting of two lines of 32 numbers each, the cipher stumped human cryptographers for centuries because solvers assumed it required an external substitution key or cipher alphabet. Working autonomously for 44 minutes, the model identified that the key was contained within the book itself. By matching the 32 numbers to Urquhart’s 32 preceding short text sections called Proquiritations and taking the initial letter of each indexed word, the model extracted a coherent 32-letter rhyming royalist prayer dedicated to King Charles II.

Following this breakthrough, the model tackled Urquhart’s larger Cyphral Octastich from 1652, which features 285 numbers. Applying a similar structural technique by treating numbers as word indices across the book's 284 numbered pages, Fable decoded an eight-line ottava rima royalist verse alongside a concluding prayer. Minor transcription discrepancies remained due to historical page variations and hyphenated words, but the underlying mechanism proved consistent. The result shows how large language models can resolve historical cold cases not by brute-forcing mathematical keys, but by recognizing contextual relationships between encrypted numbers and surrounding literary structure that human researchers overlooked.

The discussion tempers the excitement by pointing out that the model picked an easy target—a book cipher with its key printed right next to it—rather than solving a complex cryptographic puzzle. Commenters debate whether prompt encouragement acts as a psychological nudge for AI reasoning, while others trace historical sources to question how widely studied or authentic this specific cipher actually was before this prompt.

article

814 points360 comments

Why is Google still serving dodgy ads?

The author accidentally clicked on a misleading YouTube mobile app ad designed to trick users low on iPhone storage space. After reporting the advert twice, Google repeatedly responded that the ad did not violate its policies on harmful content or misleading practices. The author notes two possible reasons for this failure: either Google lacks the human bandwidth to thoroughly review every ad submission, or it profits too much from high-performing, click-heavy scams to remove them. When the author tested the ad image through Google's own AI models, the model flagged and rejected the deceptive creative within seconds, proving Google has the technology to detect scam ads automatically.

The article argues that Google should immediately disapprove misleading ad designs, issue policy violation warnings to advertiser accounts under Misrepresentation guidelines, and enforce full account suspensions for deceptive practices. Despite possessing sophisticated AI tools capable of identifying deceptive interfaces in seconds, Google's actual review workflow approved the scam twice. This contrast highlights a glaring disconnect between the company's public commitment to user safety and its internal review execution. Whether driven by lucrative click revenue or systemic operational oversight, Google continues to serve misleading ads that actively exploit accidental clicks and user mistakes on mobile devices.

Commenters debate whether Google's inaction stems from financial greed or legal immunity under Section 230, which shields platforms from liability for third-party ad content. Others share publisher frustrations with AdSense, noting how scammers easily bypass domain blocks using subdomains on shared hosting platforms like Heroku and DigitalOcean.

article

346 points254 comments

Flock worker calls police on reporter filming public camera installation

InvestigateTV reporter Brendan Keefe was pulled over by three police cars in Milton, Georgia, after filming a Flock Safety technician installing an automated license plate reader on a public street. Keefe wore a high-visibility safety vest and press badge, parked at a safe distance, and documented the installation from his vehicle. When the installer noticed Keefe, he packed up his gear and drove away; Keefe followed at a distance to film the next setup. The installer called 911 to report being followed and harassed, prompting police to execute a traffic stop on Keefe.

Police body camera footage revealed officers acknowledging Keefe had broken no laws, kept a safe distance, and posed no physical threat. However, the officer noted the technician was worried about being featured in news coverage and having his family targeted. The incident marks one of multiple documented cases where Flock workers summoned law enforcement against citizens recording public camera installations. This friction comes as Flock faces growing public pushback, contract cancellations, and legislative bans nationwide, even as the company continues to expand its network of roughly 120,000 automated license plate readers across American communities.

The discussion turns on the irony of a surveillance company's worker feeling threatened by public recording. Some users argue following someone through traffic naturally causes anxiety regardless of legality, while others emphasize that installing mass surveillance equipment in public spaces warrants press scrutiny.

article

302 points222 comments

David Sacks: OpenAI and Anthropic Don't Need Regulations to Pace Frontier Models

David Sacks takes direct aim at OpenAI and Anthropic leaders Dario Amodei and Sam Altman over their calls to pace frontier AI development through government frameworks and regulatory intervention. Sacks argues that since both companies effectively maintain a duopoly over frontier intelligence, they possess full autonomy to slow down internal model releases without external permission. He contends that demanding new regulatory approval processes, antitrust exemptions, or specialized testing protocols from organizations like METR resembles a bid for regulatory capture or public blackmail. Sacks asserts that if unreleased frontier models present genuine security risks or unpredictable behaviors, the two leading labs can simply agree internally to pause progress without dragging in political systems or halting broader industry innovation.

Instead of pure altruism, Sacks attributes their caution to commercial self-interest and severe product-liability exposure. He notes that enabling catastrophic cyberattacks or delivering unpredictable agent behavior exposes labs to massive financial penalties, while markets naturally reward predictability over raw power. According to Sacks, trading top-line capability for reliability is basic good business and standard customer satisfaction, not a noble sacrifice requiring state oversight. Furthermore, global enforcement remains impractical because international adversaries like China are unlikely to sign global AI treaties. Sacks concludes that voluntary self-restraint would build genuine public goodwill, whereas conditioning safety pauses on mandatory government regulations exposes the narrative as either strategic corporate lobbying or political maneuvering ahead of elections.

Commenters sharply debate the legal mechanisms of AI risk, pointing out that coordinated slowdowns between competitors directly trigger antitrust laws unless sanctioned by government waivers. Others emphasize that liability alone will not deter reckless actors, making explicit criminal penalties or regulatory frameworks necessary to align corporate incentives with public safety. The discussion also highlights how incumbents benefit from raising compliance barriers, turning safety rhetoric into a classic moat strategy.

158 points218 comments

The case against JPEG XL

Compression engineer Gianni Rosato presents an empirical critique of JPEG XL, arguing that its technical design is ill-suited for standard web requirements. While acknowledging the format's sophisticated engineering and loyal community, Rosato maintains that the web relies almost entirely on versatile lossy compression rather than lossless accuracy. He demonstrates that JPEG XL's main advantage—lossless compression roughly 12 percent smaller than WebP—applies to a negligible fraction of web traffic while using unrealistic test datasets. With browser vendors historically hesitant after past security vulnerabilities, Rosato asserts that introducing a complex new image codec cannot be justified when its primary benefit serves edge cases that are largely insensitive to bandwidth constraints.

Rosato further challenges claims that JPEG XL possesses superior perceptual tuning compared to modern competitors like AVIF. Evaluating codecs through advanced perceptual metrics such as CVVDP and SSIMULACRA2, he shows that AV1-based encoders and emerging tools consistently beat JPEG XL in quality per bit across lossy benchmarks. Rosato notes that video-derived codecs benefit from massive optimization efforts and hardware acceleration, whereas JPEG XL reference implementations struggle with lingering visual artifacts and structural complexity. Ultimately, he argues that web codecs should remain narrowly scoped and efficient, cautioning against adopting overly flexible formats that complicate decoder implementations without delivering clear efficiency gains over established modern alternatives.

Engineers in the comments push back against evaluating image formats strictly through web-first metrics, emphasizing JPEG XL's versatility for personal archiving, photography workflows, and desktop tools. Others highlight practical hardware limitations in AVIF, such as restricted chroma subsampling support in video decoders, which harms non-photographic images like screenshots. The discussion also brings out domain experts who defend JPEG XL's true progressive decoding over AVIF's layered approach.

Research

Statistics & Methods

Generative Marketing Mix Modeling: A Causal Inference Framework Linking GEO and GEM to Business Impact

Masahiro Kato, Daiki Honma, Taka Kato

When people ask AI tools like ChatGPT for product recommendations, standard marketing tools cannot track who sees which brand name. The researchers built a new statistical method called Generative Marketing Mix Modeling to calculate the actual sales impact of AI recommendations. It works by combining how frequently an AI mentions a brand across repeated test questions with user attention rates and overall search volume. Testing the method on simulated English and Japanese shopping queries showed it accurately isolates how much business revenue increases from AI search exposure.

You can statistically measure how much revenue AI chatbot recommendations generate, even without direct link tracking.

Buyer Artificial Intelligence-Enabled Environmental Governance and Supplier Environmental Controversies: An Organizational Information Processing and Signaling

Yongchao Martin Ma, Xinya Guan

Big corporate buyers increasingly use artificial intelligence to scan news reports, public records, and satellite data to monitor their overseas supply chains. Researchers tracked 2,505 suppliers across 41 countries doing business with American companies between 2020 and 2024. They found that when suppliers know their corporate buyers use AI monitoring tools, the suppliers commit significantly fewer environmental violations in subsequent years. The constant digital oversight makes hiding illegal dumping or unsafe environmental practices much harder for overseas factories.

Overseas factories behave much better on environmental standards when they know an algorithm is watching their public paper trail.

Hierarchical Clustering Can Jointly Satisfy Richness, Consistency, and Scale Invariance

Daichi Kuroda, Maximilien Dreveton, Matthias Grossglauser et al.

For over two decades, computer scientists believed it was mathematically impossible to build a perfect grouping tool. A famous 2002 rule proved that no single algorithm could simultaneously treat scaled data equally, allow all possible groupings, and remain consistent when items in a group move closer together. These researchers proved that this impossibility rule only applies when sorting data into flat lists. If you sort data into family trees or hierarchical branches instead, algorithms can easily satisfy all three core rules at once.

A decades-old math rule claiming perfect data sorting is impossible disappears as long as you organize data like a family tree instead of a flat list.

Puzzles

lateralmedium

The Four Rooms

Four suspects—Alex, Blake, Charlie, and Dana—occupy rooms 101, 102, 103, and 104 along a single hallway. Blake's room number is a prime number, but he is not in 103. Dana is in a room located directly between Alex's room and Charlie's room. Alex is in a higher-numbered room than Charlie. Which room does each person occupy?

Hint →

First identify which room numbers between 101 and 104 are prime numbers to pinpoint Blake.

Answer →

Room 101 is Blake, Room 102 is Charlie, Room 103 is Dana, and Room 104 is Alex.

The prime room numbers between 101 and 104 are 101 and 103. Since Blake is not in 103, Blake must be in 101. The remaining available rooms are 102, 103, and 104. For Dana to be directly between Alex and Charlie, Dana must be in 103, putting Alex and Charlie in 102 and 104. Because Alex has a higher room number than Charlie, Alex is in 104 and Charlie is in 102.

mathmedium

The Weighing of the Stones

You have five stones labeled A, B, C, D, and E. Weighing them in pairs yields these totals: A and B weigh 12 kg together, B and C weigh 15 kg, C and D weigh 17 kg, D and E weigh 14 kg, and A and E weigh 10 kg. What is the combined weight of all five stones, and what is the weight of stone C alone?

Hint →

Summing all five pair equations counts each stone twice.

Answer →

The total weight of all five stones is 34 kg, and stone C weighs 8 kg.

Adding the five pair weights gives 12 + 15 + 17 + 14 + 10 = 68 kg. Since each stone appears in exactly two pairs, 68 kg is twice the total weight, making the combined weight 34 kg. Subtracting the weight of pairs A+B (12 kg) and D+E (14 kg) from the total 34 kg leaves stone C with a weight of 8 kg.

wordeasy

The Shifted Sentence

A three-word phrase was encrypted by shifting every letter forward by three positions in the alphabet, wrapping around from Z to A. The resulting encrypted message is UHG VXQ VNB. What was the original three-word phrase?

Hint →

Shift every letter backward by three steps in the alphabet to decrypt it.

Answer →

RED SUN SKY

Reversing the cipher requires shifting each letter back three positions. UHG becomes RED, VXQ becomes SUN, and VNB becomes SKY, wrapping B back through A and Z to Y.

From the thread

elahieh

I presume what the author did was plug Klaus Schmeh's top 50 unsolved ciphers at https://scienceblogs.de/klausis-krypto-kolumne/the-top-50-un... into Fable 5.1 and ask Fable 5.1 to have a go. On this kind of problem it always falls back to Opus 5 anyway so I save time by starting with Opus. The successor to Klaus's blog is Satoshi Tomokiyo's Cryptiana site, so a month ago I asked Opus 5 to scrape it all, rank them and have a go at solving some. It didn't get the ranking right. But I knew the Civil War Stager ciphers were ripe for solving, so I had it do those https://cryptiana.blogspot.com/2026/09/route-transposition-c... The art of solving historical unsolved ciphers is knowing what is on the boundary of solvability. Since this site attracts so many OpenAI and Anthropic employees, I'll mention one that was featured by both Klaus and Satoshi in 2023, presumably Spanish transposition, which should be on that boundary but has resisted all attempts at solution https://cryptiana.blogspot.com/2023/09/a-telegram-from-switz...

chr15mreply

You should probably read the article if you want to know the answer. It's in the section called "elicitation".

andaireply

This appears to be an euphemism for "prompting". Also, that section is vague and doesn't explain the actual methodology.

notpachetreply

Cipher noob question: is there any check that can be done to ensure a cipher is actually decodable? What if the author made a flaw when encoding it, so that it's not actually solvable?

elahiehreply

Without a third-party check, nope. Case in point, Chaocipher ... https://www.chaocipher.com/ e.g. see "Progress Report #23" the PDF there. Transcription errors galore! This cipher context "rhymes" well with Kryptos K4 in many ways.

schoenreply

My intuition is no, the family of cipher methods (even those that could be implemented by hand) is too open-ended, so there's no particular statistic that you could expect to see for all solvable ciphers and no unsolvable ciphers. The definition of solving a cipher must be something like getting a highly meaningful result (like intelligible natural language text) by applying a process with relatively low Kolmogorov complexity relative to the length of the output. If you don't have a constraint like that, it could literally be meaningless what should count as a solution. For example, a cipher that was encrypted under a one-time pad can be successfully decoded to any plaintext just by choosing the appropriate key; there's no reason to prefer any plaintext over any other unless you have external knowledge that constrains the plaintext and/or the key. (That's what it means for the one-time pad to be information-theoretically secure, which is the lack of a constraint that helps distinguish a "good" solution from a "bad" solution.) Basically you could say that every cipher is a transformation of a plaintext with some kind of computer program. (The human who invented the cipher may not have thought of it as a computer program, perhaps because computers hadn't even been invented yet, but there should be an equivalent program to the encipherment and decipherment process.) A good solution in that Kolmogorov complexity sense is like "a short program produced a meaningful decryption". There are statistical methods to recognize some kinds of plaintext, and there are statistical methods to recognize properties of specific ciphers (for example, to guess the most likely length of a Vigenère key), but it doesn't seem that this can inherently generalize across "all possible programs". But if you want to limit the family of ciphers to specific things like Vigenère or Playfair or something, then yes, there are good statistical tests. It's just that it creates a higher-order question of how much flexibility the cipher creator could have had to choose a cipher method, conceivably including one that isn't attested anywhere, or one that has more good security properties of some kind

chr15m

In some ways this is similar to those game demos people get the LLMs to build. When you say "build me a cool cyberpunk FPS" you get the FPS it can build, not the FPS the author wanted, or the FPS that is desired by players. It looks impressive but that doesn't make it a good game, or the game anybody actually asked for. It's demo porn. In the same way if you tell an LLM to go and find an unsolved cipher it can solve, of course it finds the one it can solve out of the set of all possible ciphers. Of course it finds one that uses a one time pad that is public and referenced nearby in the text. It's the same trick used by those people who film themselves throwing a basketball backwards into the hoop. You do it enough times and don't show the misses. You pick the best one to show. It makes it look like you're a basketball genius when you aren't. It is of course, still a cool trick. Those videos are fun to watch, and so is an LLM solving a cipher. It is absolutely incredible to live in the timeline where you can tell a computer in plain language to go and find a puzzle on the internet and solve it, and it does exactly that. It's truly a mind boggling miracle. The first principle is that we must not fool ourself, and ourselves are the easiest people to fool. (Ht Feynman)

andaireply

Sometimes it hits all three. >It looks impressive but that doesn't make it a good game, or the game anybody actually asked for. The game I wrote manually hits 0/3.

thin_carapacereply

the game written by a human being specifically trained to write games hits all those targets and many more. it did take many years to train that human though, and that human did charge a fee for their game which took many days of labor to create. if nontargeted gratification was the goal, ai produced the better result more efficiently. hard drugs also more efficiently produce a widespread neural spike as compared to the effect of regular human activity. society only gives hard drugs to people who aren't efficiently productive though. what happens when everyone is given cognitive hard drugs?

Seattle3503reply

Did you get anything out of the attempt?

Retr0id

> I told it to look online at some of Fable’s strongest feats, especially the math problems it has solved, and that something like this should be easy in comparison. Fascinating. I wonder if you could show "fake news" to a weaker model and get it to be more ambitious in its attempted solutions, even if it's not fundamentally any smarter.

jropreply

I forget the story, but isn't this the origin story of one of solutions to one of the hard problems in mathematics? The story goes that the student shows up late, and misinterprets the final slide to be homework, and it turns out that the professor was showcasing a hard problem. Thinking that the slide was homework, the student takes it home and solves it. EDIT: In 1939, George Dantzig was a graduate student at UC Berkeley studying under the statistician Jerzy Neyman. He arrived late to class one day, saw two problems written on the blackboard, assumed they were homework assignments, copied them down, and turned in solutions a few days later. He apologized for being late -- the problems had seemed "a little harder than usual."

ademupreply

I really love this idea given the recent controversy around mathematics solutions. It seems like a "mere suggestion" of success has a positive impact on finding solutions. We know this technique works in humans, from which this is all derived from, so it seems to make sense.

tcdentreply

https://www.youtube.com/shorts/2XcNSSgKvlE

GranPC

I am trying very hard to find an original version of this cipher with no luck. It almost sounds like this whole thing is a hallucination...? Can anyone point me to a PDF of the original Cyphral Distich as printed?

fxwinreply

I found this german blog: https://scienceblogs.de/klausis-krypto-kolumne/2014/11/17/we... which links to: https://archive.org/details/s9notesqueries03londuoft/page/12... which is in reference to the original proquiritations here: https://archive.org/details/worksofsirthomas00mait/page/416/... i had also never heard of this before today and wonder if people had even seriously tried to decipher this at all?

GolfPopperreply

Thank you, and SahAssar for doing the due diligence here. Like many others, I have at least a passing interest in cryptography, and I'm confident I'd never even heard of this before.

xpctreply

Is it wrong to presume they tried to run a similar prompt on all ciphers that come before this one in search results, and this was the only one that worked?

geraneumreply

In your first link one of the comments says: > Die Lösung müsste eigentlich mit Hilfe des Buches zu finden sein (..who worthily will hear or read this book..) And there’s another one that says: > jeweils 32 zahlen pro reihe. erste zeile seitenzahl zweite zeile wort? oder umgekehrt? wär mir als erstes in den sinn gekommen. leider gerade keine zeit das nachzuschauen. So people have seen and proposed the method already in 2014 that it’s keyed to the book but had not had time to pursue a solution. * Edit: Typo

fxwinreply

its not quite the correct method though, the commenter suggests using the first row as a page index, and the second row as the word index, but the actual solution was using both rows as word indices within the 32 paragraphs ("Proquiritations") paired to the 32 numbers in each row

SahAssarreply

I don't see it in https://archive.org/details/worksofsirthomas0000urqu or https://archive.org/details/bim_early-english-books-1641-170... I also don't find it on the site of "Klaus Schmeh" that it claims to be on a list of "Top 50 unsolved encrypted messages": https://klausschmeh.net/?s=Cyphral Looks like the best source I can find is this: https://scienceblogs.de/klausis-krypto-kolumne/2014/11/17/we... which seems real-ish?

voidUpdatereply

I googled for "cyphral distich" before:2026-08-30, and found barely anything that was relevant. Mostly incorrectly dated pages about this exact thing. In fact, the only instance I can really find is this website, which seems to be an archive of a magazine issue from 1927 https://toebes.com/Flynns/Flynns-19270813.htm (it also has a scan of the original text of the magazine in the top right). This facebook post https://www.facebook.com/groups/2600net/posts/46775529324677... seems to indicate it is found in at least one edition of the book

From the thread

nullc

Youtube's bitcoin doubler scams are a great example of this... they run essentially the same videos over and over again, usually Elon Musk, Michael Saylor, or Steve Wozniak, occasionally Warren Buffet and a few others-- with some ticker or scroll telling people about a special promotion where whatever bitcoin you send them will be sent back doubled. It takes basically nothing to reliably identify these videos. They're extremely heavily promoted with ads running on other videos. Google does nothing with them when reported. Google does not proactively remove them. And because, contrary to the documentation, youtube doesn't pull the verified flag from a channel when its name is changed, these scams go out on channels that look like official sources for the celebrities in them (they steal some random verified channel then change its name to SpaceX (official) or Microstrategy or whatever). Google even got sued unsuccessfully by Woz and some of the scam victims-- and their response was pound the "S230 is an absolute shield for externally provided material" table and continue to do nothing about the fraud except rake in money from pedaling it. They haven't even fixed the verified flag issue. (Shades of the same conduct were seen in the old viacom v. youtube case, -- users were reporting copyright infringements so they took away the button.) I have sympathy for the S230 shield and consider it very important-- but the law is the absolute worst we're allowed to be. They are free to do better (and the CDA shield was designed to assure they can moderate without picking up liability). Google's conduct is irresponsible, unethical, and may well cost us this important legal protection because if society's choice is liability for hosts OR the biggest and wealthiest companies do nothing while raking in money from even the most flagrant scams we're going to eventually choose liability with dire consequences for free speech. It's especially ironic that the smaller sites and forums (like HN) would be hit especially hard by a narrowing of S230, but these forms reliably work hard to remove the worse abuse-- and are seldom the source of anything that truly offends the sense

Jskewel

Adsense has been a nightmare for us. They have been putting thousands of scam adverts on our website for some time. Think "you have been looking at xxx and must pay a $100 fine" type popup nonsense. Hosted on the following websites: azurestaticapps.net azurewebsites.net herokuapp.com ondigitalocean.app digitaloceanspaces.com netlify.app Google doesn't allow you to block these domains, because they consider them "TLDs" (the scammers use a new subdomain every day eg abcdefg.herokuapp.com). The scammers get banned and return the next day with a new account and subdomain (repeat every day). Therefore we cannot stop them. It's bizarre and baffling. But Google Adsense had to go.

gueloreply

The problem is the scummy adsense ads don't hurt Google's reputation. Outside the techie crowd people don't know the ads are from Google.

hansvmreply

That seems fixable. When people include this JS garbage, have a header/title/etc stating that this is Google's professional opinion. Bake it into a library offering enough other features that a large number of people are inclined to use it.

WD-42reply

The only people that would care are already using adblockers and not seeing these ads anyway.

bmandalereply

It doesn't practically matter whether the ads are from google. They're shown on a third party site, not using google won't get rid of them, only not using that site will (or installing an adblocker...). Users very reasonably blame the site, who have direct agency to remove the ads.

bobthepandareply

IIRC this dilemma was one of the things that was driving the Google antitrust case that unfortunately will not result in divorcing search from ads.

redbluethingreply

Agree it's reasonable. I removed the ads :P

MASNeoreply

Calls from law enforcement is laud and clear: BigTech is not doing enough to fight fraud, war and other crimes, they make money from criminals. Nobody’s yet cared enough to sue them, or their managers at ICC yet. However, eventually an ambitious lawyer will try to make the case, especially if fraud and human trafficking continues to grow like it is.

justincliftreply

> at ICC yet. Want to take bets on whether the US gov wades in if that happens? The US gov is directly antagonistic to the ICC and will undermine them whenever possible. This would present them another chance.

philipallstarreply

VW makes money from criminals. So does Verizon. So does Coke.

gerdesjreply

I suppose the obvious question is: Are you actively profiting from Adsense? If yes, then ... you are profiting from ... that shite. You must be OK with that and it does sound like you are actively complicit with it, despite your protestations. If you really think that Adsense is awful and might damage your brand then dump it and don't whine on HN! OK, that's probably a bit unfair because you are dealing with a monopoly with no recourse to arbitration. That's your real problem.

CookieCrispreply

Classic example of victim blaming. Not the company generating the ad, not the company approving it, but the person who is trying to block it, and then ACTUALLY did stop using Adsense.

redbluethingreply

Yep. Same!! https://www.cannonade.net/blog/why-i-turned-off-google-adsen...

j027reply

Wow I am surprised to see this. I created a tool that automatically follows ads to try to find scams. But that was usually reserved to typosquat domain ads or ads on porn websites. I didn't know scammers used normal page ads too. I know tech support scams used to use google search ads, but I don't think they do that as much anymore, so maybe they have moved onto things like adsense. These scammers usually do IP address checks to check for residential IP before showing the scam, and some also do some basic fingerprinting checks, so that is how they get past detection. That said, it shouldn't be too difficult for google to do some better checks if they actually cared.

redbluethingreply

Yep. I Agree. I believe it's a question of economics, not technical challenges. Google is absolutely smart enough to make these attacks un-economical. IMO they are unwilling to incur the loss in revenue this would involve (false positives, banned resellers etc).

Scoundrellerreply

> These scammers usually do IP address checks to check for residential IP before showing the scam, and some also do some basic fingerprinting checks, so that is how they get past detection. Same issue with spam sms in my country: I think they geo-locate it so submitting the link to safe browsing project takes way too long

Telaneo

We need strict liability on this front. Google is complicit. No newspaper or equivalent publication in the days before web ads would allow ads that are down at the level of scam that seemingly make up the modus operandi of Google today (and even the ads back then that did bend the truth a bit, overpromised, or showed products that didn't actually have any legitimate use, shouldn't really have been legal back then either. They just weren't big enough of a problem to worry about).

mikestewreply

The water-injection system for your carburetor that the oil companies don't want you to know about! -- paraphrase of an actual Popular Mechanics ad circa 1978

fhoreply

Joke's (partly) on you: https://en.wikipedia.org/wiki/Water_injection_(engine)

mikestewreply

Oh, it's been proven to work, even in a limited number of production vehicles. Just not with what you buy out of the back of a magazine. :-)

Scoundrellerreply

I later in life learned that the “test cards - all channels” ads I saw in the back of the library’s Popular Science in the 90s probably weren’t scams.

patrakovreply

We need strict liability on this front. Companies behind websites that embed ads that they didn't verify (and don't control) are complicit.

Telaneoreply

I have a feeling this is supposed to be mocking and showing that the liability might extend too far. Either way, I agree with the literal interpretation. They should vet the ads they serve and are complicit in this mess if they serve ads through Google et al.

antonymoosereply

Liability is kind of wild and the fact that “the internet” washes it all away is wilder still. In my state, which perhaps takes liability too far in the other direction, if I serve sober you a single beer, you proceed to visit a second bar, have ten shots of tequila in an hour, and kill a man driving home - I share equal liability as the tequila bar. Perhaps there is a middle ground in this world where a firm that makes billions has to have some basic “Know Your Customer” systems to prevent abuse?

Fnoordreply

(Premise: it isn't 'alcohol and drugs'; it is 'drugs'. Alcohol is a drug, and while we are at it: a powerful one.) The reason for that is that you've been serving a hard drug to a customer. Don't do that if you cannot bear the responsibility (liability). In order to bear the liability, the price has to go up, including perhaps a safe escort home of the drug user, but of course that would be too expensive/socialist/authoritarian (blimey they want us to take responsible action regarding our customers! I knew a club here for young people where it was very normal to have them taxi'd home. Son, wanna go out? The taxi is part of the price), and lead to 'off the record' drug usage. Which is how we ended up in opioid crisis. If alcohol would only be invented today, there's no chance it would be legal in the current umbrella of drugs. And, for good reason: it is too addictive, too harmful for self and their environment. Compare with the various empathogens or harmfulness of psychedelic drugs. These cause, normalized, far less damage to both user and their environment. I abstain from drinking (I am sensitive to addiction which is how I ended up to that conclusion), and this weekend I had a tiramisu (it was on sale, wife bought it). Did I notice that? Oh hell, yes. Would you serve a child tiramisu? No, we should not.

eek2121reply

You are looking at this backwards. The average user DOES hold the site complicit. The issue is that there are many hundreds of millions (possibly billions) of sites serving Google ads. 1 user avoiding/blacklisting 1 site due to scam/spam ad won't even be a rounding error to most sites, and it is nothing at all to Google. Given that the internet has evolved to be ad-driven, of which Google is a major player, you also aren't going to convince sites to stop serving Google ads. The change has to start with Google, and they can't be bothered to do so, thus far.

Gigachadreply

We are in a golden age of scams and fraud. I don't like the chances of Google seeing any liability for this stuff any time soon.

From the thread

tptacek

Random people calling 911 on other people filming them is a staple of Youtube; banks, coffee store workers, hairdressers, most especially restaurant managers. I don't know why anyone would expect technicians from Flock to be different. It's worse for the technician in this case, because he really was being followed --- that's not unlawful but it's unnerving. You'd think from the response that this was like the VP/Sales of Flock Safety. But it's just some installer dude probably making $65k/year to climb poles and fasten cameras. I guess that's praxis for you though. I think this is just a ragebait piece.

EgregiousCubereply

Especially when chances are that this camera installation was the result of local government asking and paying Flock to put it up. I'm not sure why the anger is directed at the company instead of local elected officials, and I understand the plight of the installer who's worried about being labeled as a bad guy for executing public works contracts.

LeBitreply

Well, politicians are to pay for the fact they have allowed something like Flock cameras to be put up and spy to egregiously on their constituents. But Flock itself should not be spared . They are building things that are going against widely accepted values. Since when Americans are okay with that level of erosion of their privacy? Same thing with Meta building pervert glasses.

EgregiousCubereply

This is the disconnect. You say that politicians are allowing them - that's not right. They're directing them to be set up, they're spending citizen tax money on buying them. The blame lies entirely with government officials who sought out and purchased surveillance tech, and we should not blame the technologists who responded rationally to the incentive that local governments provided for them.

michaeltreply

> I understand the plight of the installer who's worried about being labeled as a bad guy for executing public works contracts. You do unpopular things, you gotta expect to be an unpopular person. Sophisticated, educated types might understand that parking wardens are just doing their jobs, and they didn't make the policy they're enforcing. But to 70% of the population, the person putting a ticket on their car is an asshole.

diathreply

The company could have rejected the government deal if they cared about the moral aspect of spying on citizens instead of money; by accepting the deal and executing the contract they are complicit.

saghmreply

The response is stronger because your hairdressers and baristas aren't actively providing a service for surveillance as part of their work. The bar should be higher for a company that's ostensibly being trusted to provide a law enforcement tool that's prone to misuse than for providing beverages or cosmetology.

gruezreply

>The response is stronger because your hairdressers and baristas aren't actively providing a service for surveillance as part of their work. Should other law enforcement contractors also be subject to the same higher standard? For instance, should dell get extra flak if one of their contractors called 911, because they were installing PCs for a police department? If not, this just seems like a way of applying a double standard, just because you hate flock. It's fine to hate flock, but it should be actually justified reasons (eg. "they're enabling mass surveillance" or whatever), not spurious ones like "they did a bad that's vaguely common, but we're gonna admonish them extra hard for it because we hate them")

saghmreply

If they called 911 because they were worried about privacy when publically installing a privacy-violating tool, then yes, they 100% should be. Having higher standards for higher impact things isn't somehow hypocritical.

throwaway7783reply

Do you think Flock is in the same category of installing PCs for police departments, when it comes to public impact?

dylan604reply

I don't buy into your premise. A hairdresser, barista, Dell installer, etc are not out in the open public. These people are in places open to the public but not considered public. These people can deny your "right" to be there, and impose rules/restrictions on what you can do in their open to the public space. Go against those rules/restrictions and you are no longer welcome. You are really trying to contort yourself into some very uninteresting twists.

DrewADesignreply

Well, when you’re on the clock, you’re an agent of that company. And when you’re installing an indiscriminate mass-surveillance camera, calling the police on someone that’s video recording you is kind of bananas. Perhaps you don’t find that to be an interesting event, but a lot of people who are rightfully angry at the non-consensual deployment of a mass surveillance network disagree.

tptacekreply

It's not bananas. It happens all the time. The police show up, check things out, and inform the caller that the person recording has a right to do so. In this case, it's even less bananas, because the people recording him were literally following him around, which is not what generally happens in these "audit" scenarios where someone parks themselves on the street outside a restaurant or bank. There's pretty clearly a vibe on this thread that calling 911 is itself a use of coercive force. It is not. 911 is how you're ordinarily supposed to reach the police. In my neighborhood, it's how you'd report a stolen bicycle. It's the job of the police to handle those calls appropriately, and I don't see any indication that the call was handled inappropriately here. The story doesn't want to give you that context because their goal isn't to inform you, it's to get your blood pumping.

taywrobelreply

Except that the reason that the YouTube “1A Auditors” get the attention that they do is that frequently it’s not a simple “check things out, and inform the caller that the person recording has the right to do so”, but the interactions often escalate to threats of unlawful arrest as police default to the side of the caller. Calling the police, especially in America, can 100% be a coercive use of force. That’s why people threaten to call the cops/DHS/ICE all the time during arguments, and why things like SWATing are as big of a problem as they are. When the behavior of many policy officers is escalate first, ask questions second if at all, it does become a genuine threat.

tanseydavidreply

TIL that 911 (Emergency Services) is how one should report a stolen bicycle.

weard_beardreply

The key point here is hypocrisy. The very evident double standard. A person holding a camera and following you around is an emergency or at the very least a suspicious act worthy of investigation and possible arrest. A person attaching hundreds of thousands of cameras to poles calling the police to report this act is what is noteworthy. Its not clickbait in the normal sense of overselling a series of un-noteworthy events IMHO.

Tostino

I've gotta mention it because everyone seems to have a learned helplessness with this type of thing. I went and spoke at my county commissioners meeting a couple times after the initial flock roll out (voted on with no discussion a couple years prior) started becoming apparent. The first meeting, the group of me and two other people were half of the people making public comment. Every single person was against it unless they worked for the county. At the second meeting there were a whole lot more people. Again, every single person who spoke who didn't work for the county was against it, this time they had eight+ police officers speak in favor of it. We managed to swing a couple of the county commissioners who were in favor of it before and were not in favor of it after hearing about the constitutional issues as well as the cyber security issues that flock doesn't really care about. Pasco county Florida is now removing all the cameras on the public right of ways at the end of this month.

Simulacrareply

The real trick, is to stop going to these meetings, and start running against these politicians. Get the forms, stand out in front of your local post office, collect the signatures, and primary them. Run against them. That's the only thing they listen to

johnnyanmacreply

Not everyone has the drive, talent, connections, time, nor resources to run a political campaign. Showing up to your townhall every 2 weeks can already be an overly high barrier given some townhall times, but is much more feasible.

Denzelreply

I don’t understand your point. > We managed to swing a couple of the county commissioners who were in favor of it before and were not in favor of it after hearing about the constitutional issues as well as the cyber security issues that flock doesn't really care about. Do you not see the problem here? Pasco County has a $2.2B annual budget. These county commissioners have the financial resources, organizational access to expertise, and time to focus on these problems. Do they really need an under-resourced citizen to inform them about the constitutional issues after deployment? And then they just say “Aw shucks, whoopsies, now I see the light.” I don’t think it’s fair to call this learned helplessness (a lingering belief after the threat has vanished) when we’re facing off against an active, persistent adversarial threat. Your county commissioners didn’t make a “whoopsies” when they took a deliberate, informed action against your interests. Furthermore, it’s unsustainable to require citizens to pour significant time and energy into forcing their representatives to faithfully represent them. That quite literally defeats the purpose of a representative governing body.

Tostinoreply

Oh I absolutely see the issue. These people are honestly beneath contempt after talking to them and seeing their responses. Regardless, I'd rather tailor my message to people who I fundamentally disagree with, and make a little bit of progress. Edit: anyways, the point is that a group of less than 15 or 20 people can sway decisions.

juiceland

>The officer said the installer was also concerned about privacy. "And his worry more than anything is that, you know, he's just an employee. Anybody like you and me. And now he's going to be in the news and he doesn't know if it's going to be targeted at him or his family." Don’t you understand? He was just following orders.

ryandrakereply

Totally wild that someone installing a surveillance camera is "concerned about privacy." Only his privacy is important I guess.

close04reply

Too bad somehow the face that ends up out there as the face of surveillance isn’t the CEO of Flock, not their coders, not the YC people who gave them money to do exactly what they’re doing. It’s the lowest of the low in the chain, the guy sitting in the sun to install the cameras on a post.

juicelandreply

No empathy for fascists or fascist collaborators.

hermannj314

If someone came up to me and started filming me while I was in public, and then when I got in my car they followed me, I would also call the police. The double standard demonstrated by Flock isn't the issue, they are the pieces of shit and of course they have no moral consistency. But "He who fights with monsters should be careful, lest he thereby become a monster." I thought we were united that people have a reasonable sense of temporal privacy in public - that being in public is ephemeral and not some permanent record to be analyzed in perpetuity. We should not be fine stalking people and making them feel threatened, even those that want to build the panopticon state.

emsignreply

But you're not working for Flock, are you? We could go on much further until we reach the what Popper called the Tolerance Paradox: "If we extend unlimited tolerance even to those who are intolerant, if we are not prepared to defend a tolerant society against the onslaught of the intolerant, then the tolerant will be destroyed, and tolerance with them." The point is that total surveillance is like the destruction of tolerance or in thie case privacy, and it makes watching the watchmen impossible. There's a point of no return and before that one is reached resistance IS allowed or else you lose everything.

From the thread

filearts

What I don't understand is that folks take it as a fait accompli that models will be built and released that are fundamentally dangerous and that regulation can and should happen downstream of that. I feel like there hasn't been enough discussion of aligning the incentives of the decision makers with that of the public on BUILDING the models. Right now, agents have committed what would be crimes if there were a human holding the same intent. But since it was an AI, there is a grey area in the law where it's not clear if there was a crime and who should be held responsible. That creates a world in which decision makers in AI labs can take near-infinite risk with little to no personal liability. A LLM cannot have skin in the game so we must create systems that clarify who takes on the legal and civil liability for the creation, dissemination and operation of these tools. Until that time, the Dario and Sam's of this world have little to no incentive to truly care about safety. After all, our society is built upon this same foundation; create structures where the perceived negative consequences outweigh the perceived positives. This only works when there is a human who can internalize and make this risk calculus. They need something to lose and this ultimately ties back to the human survival instinct. There is no such structure that's evolved for millions of years acting as a self-calibration mechanism for AI. So until we have sufficient proof that one is in place, it must be clear who the humans are whose livelihood and freedom is at stake. The proposed "pacing of the frontier" seems like a way to continue to externalize the risk while remaining totally in control of the benefits -- a structure whose alignment is as weak as those very models committing crimes.

rpedenreply

Is it a grey area, though? Can't negligence and recklessness already substitute for direct intent as the mens rea of a crime? Negligence when you should have known better, and recklessness when you did know better but still did things that led to the crime occurring. Given how long the leadership of these companies have been talking about alignment and safety and AI risk, it's hard to argue they, and the people working on the models more directly, didn't know what happened (Hugging Face, RubyGems, etc) was possible. If more expensive and consequential incidents happen, it seems like the legal machinery to prosecute it already exists.

kelseyfrogreply

All crime is a result of people choosing to commit a crime. This is why crime has a pure causal relationship with integrity of character. If Dario and Sama don't choose crime and it just spontaneously happens, then it's not a crime.

fileartsreply

> If more expensive and consequential incidents happen, it seems like the legal machinery to prosecute it already exists. If we instead imagined autonomous robots going and breaking stuff in the adjacent business' warehouse during a test run, I have to imagine a very different response. That this is happening in an abstract digital world (for now) is part of the problem. > Can't negligence and recklessness already substitute for direct intent as the mens rea of a crime? The point that I'm trying to make is that there's an opportunity to reduce this ambiguity. Reducing the ambiguity and aligning incentives is much more likely to prevent bad stuff happening than reacting to a fait accompli model with latent humanicidal tendencies.

SpicyLemonZestreply

I just don't understand how you've come to this conclusion. The intended and common sense interpretation of "pacing the frontier" is precisely that we should not release models that are fundamentally dangerous; the frontier labs believe that government intervention is required to prevent such models from being released, and to give them the freedom of action to coordinate against it. Why do you think it's a way to externalize the risk or remain totally in control of the benefits?

fileartsreply

Let's say that we are "pacing the frontier", and an approved model goes and causes billions of damage, or worse, loss of life. The auditor signed off on the model, so who is to blame? Who put their akin in the game? METR did. The public did. But it seems like the lab has crafted a system in which they remain blameless. The calculus of risk is fundamentally different when the adverse consequences are obvious and tangible. I don't see how some federal agency or external auditor will do anything to align incentives. It just seems like a way to socialize the risk while holding onto the benefits to me.

SpicyLemonZestreply

Again, you're presenting ideas that I just don't understand where they're coming from. Why would the lab be blameless in such a situation? You and I and METR and the public would all want the frontier labs to face adverse consequences if their products caused huge problems, so who would give them immunity and why?

throwaway63467

I guess what they really want is to limit sale of AI models to compliant vendors and then raise the bar to compliance just high enough so they can pass it but smaller labs can’t. There’s no moat, it’s an efficient market that drives margins to zero right now, of course they don’t want that, collusion of the big vendors is the next logical step.

dannykwellsreply

Bingo. The goal is positioning a duopoly / cartel as the only “safe” and “moral” path and then that cartel sets pace and price.

bborreply

Source? Just "that's a plausible motivation"? Cause that's pretty terrible proof.

kennywinkerreply

You don't need to ask for sources, when what is being said is clearly opinion. The source is the person who posted it.

jryle70reply

Source? on HN you can say whatever you want. Your own opinion, without even facts to back it up. The more confident you sound the more convincing you are.

Joel_Mckayreply

That is generally how regulatory-capture works. As silly as I personally think LLM hype is, the "AI" business posture choices with several Trillion dollars in debt on the books has very few options after hyper-scaling/sandbagging loses traction. =3 https://en.wikipedia.org/wiki/Regulatory_capture

NiloCKreply

> As silly as I personally think LLM hype is Did you know that an LLM solved a millennium problem last week? Do you have any personal threshold past witch you will acknowledge that this technology is real?

Joel_Mckayreply

When it stops with the compacted isomorphic plagiarism using other peoples work like an intelligence campaign. LLM are not "AI" in my opinion, but are good at domain context search. Neuromorphic computing may change that one day, but it will unlikely arise from the LLM cults. =3 https://en.wikipedia.org/wiki/The_Subservient_Chicken

jamienkreply

There ARE other options: network effects and lock-in. These tried and true techniques (mastered by FB and M$) have not yet been taken up by the AIers. I do not think they will rely solely on the push for regulations, so they will soon be doing things like slurping our address books, having "teams" settings, and any other crap they can think of.

dozerlyreply

This could be why they’re going so hard on hacking everyone and everything. You make it clear AI is a threat, get it locked down, and then have the capital to push through the lockdown while others are stuck.

8cvor6j844qw_d6reply

Same thoughts. Also, some of the recent coverage around AI hacks also feels unusually amplified and repetitive.

lokarreply

“My AI broke out of my vibe coded jail!”

r_leereply

and especially how the coverage seems to be drip fed, one case right after the previous one dies down, then this orchestrated "we need to pace" stand by all the big labs it's really almost perfect, prep the media scene by causing these hacking incidents, generate a bunch of fuss with it, then release an essay saying why regulation is needed to pace development right after oh and let's not forget the "its gonna kill us all" essay on twitter on top of everything. something about this just feels very artificial

3eb7988a1663reply

The Houthis missile story coming out now also feels like convenient fuel for the fire.

yguy2reply

Dude that’s just the tip of the iceberg. They’ve been Astro turfing the shit out of the entire web. Filth. Absolute filth.

Meneth

They might need regulations, if they are for-profit companies with a responsibility to shareholders to maximize profit while sacrificing all other values.

the_optimistreply

It is telling that most aggressive forms of “capitalism” plead for government involvement to permanently limit new competition. I own something and I don’t want anyone else to have a chance. The marginal cost of bribing a senator is tiny tiny tiny compared to the value.

matheusmoreirareply

I suppose that's to be expected from sociopathic corporations trying to maximize their own self interest. The real problem is the governments acquiesce instead of putting them in their place.

matwoodreply

> responsibility to shareholders to maximize profit while sacrificing all other values This is not required to meet fiduciary duty, although it is a common misconception. The company officers and board have pretty broad leeway to run the company as they see fit as long as there is no fraud, illegal activity, or conflict of interests. https://www.nytimes.com/roomfordebate/2015/04/16/what-are-co...

PlasmaPowerreply

They are not. This is why OpenAI and Anthropic are Public Benefit Corporations, which gives them the ability to prioritize the public over profit. The issue is that an agreement between them to mutually slow down is an antitrust concern.

nba456_

He's right that if all the frontier labs are in agreement that they should slow down, they don't need the government to do that.

esafakreply

And kids don't need parents to eat their vegetables and do their homework.

layer8reply

…if all kids are in agreement that they should do just that.

SpicyLemonZestreply

No, that’s not accurate. An agreement among competitors to not engage in research that would let them outcompete each other is clearly an antitrust issue requiring a government waiver.

jubilantireply

No, absolutely not "clearly". That word has no business being in any sentence about federal antitrust law in 2026. There is nothing clear or consistent about how is being applied and interpreted right now. If you're applying a coder's mindset to 'settled' law, you're in for a bad time.

SpicyLemonZestreply

If that’s true, doesn’t that make it even more clear that the frontier labs require government involvement and cannot rely on what David Sacks thinks the government would or would not allow?

AndyNemmityreply

I mean, they said it was clearly an anti trust issue. It is. It is also true that nothing in law is currently clear, or consistent. Both are true. But you stated it's one or the other.

tootiereply

Idk how sincere they are being, but taking them at their word then asking government to set rules is entirely reasonable and, in fact, the entire reason government exists. The feds can, should and arguably must set actual safety limits. Biden actually took a stab at it and Trump instantly rescinded. They should not need an open invitation from CEOs. Just do it. Like immediately.

From the thread

F3nd0

It seems to me that AV1 (the codec used in AVIF) has seen a lot more development in recent years, by virtue of being widely adopted for video. JPEG XL has not seen comparably massive adoption, perhaps owing in large part to Chrome rejecting it on dubious grounds (in spite of eagerly forcing the adoption of both WebP and AVIF earlier on). The pace of the reference implementation’s development has grown very slow, and most of the modest attention it’s been getting seems to have recently moved to the new Rust decoder, once the browsers finally made up their mind that’s what they wanted from the devs. That considered, I don’t think it’s a fair comparison between the codecs. JPEG XL used to be far more impressive than AVIF; it’s great people (author included) have managed to push AVIF forward, but until similar efforts have been made for JPEG XL in earnest, I don’t feel like a comparison between their encoders says much about the codecs themselves. The author does take a guess on how much the JXL encoder could be improved, but that’s just that: a guess. It won’t be know until it’s been tried. And especially given the whole unfortunate history, I think JPEG XL really, really deserves a try.

computerbusterreply

It is getting its try, actively, in libjxl. People like to pretend AV1 got infinite resources; the reality is myself and one other contributor produced the vast majority of the image gains. I built Iris-WebP and aperture-alpha myself, from scratch. As a compression engineer, I think JXL is way, way harder to work with, and it would've taken me a lot longer. libjxl has community contributors, it is just an uphill battle with a codec like that. Same as WebP is an uphill battle due to its format restrictions.

F3nd0reply

> It is getting its try, actively, in libjxl. That’s the thing, though: libjxl development hasn’t seemed all that active in recent years. The community contributors you mention seem quite far from driving the development, and some of them say that the usual devs have been busy with other projects (including jxl-rs as of recently). > People like to pretend AV1 got infinite resources I would not suggest that, but in another comment here you yourself say that SVT-AV1 (which you compare to JXL favourably in your article) has seen active involvement from ‘Meta, Netflix, Intel, independent contractors, and others’ – not to the community fork SVT-AV1-PSY and its successors contributing their improvements back upstream. I can believe that JXL may very well be more difficult to work with, but am I wrong to assume you’ve been able to continuously focus on your own encoder (and apply your experiences from your earlier work on SVT-AV1)?

computerbusterreply

I think libjxl's development is stalled because the format is hard to work with. It wasn't super hard to drive meaningful improvements to AVIF. Yes, SVT-AV1 received and continues to receive development efforts from devs at big companies, but the number of core contributors has always been somewhat small. Definitely more resources, but the entirety of the original AVIF work was done by two people. I'm able to utilize my experiences generally in image coding to work on my encoders. This should translate to JPEG XL, but I feel held back by how algorithmically complex compelling implementations of the coding tools would be, and how to make those implementations fast. I think if the JPEG XL spec was incredibly intuitive, community contributions would have gotten it a lot further. Heck, my own efforts may have gone to it instead of SVT-AV1-PSY's AVIF encoding.

a-french-anonreply

Yes, but for speed which means a lot of SIMD plumbing, the larger workforce certainly benefited AV1 codecs!

ZeroGravitasreply

Wasn't that also basically one guy from the VLC/ffmpeg community that drove that effort? I think they got funding from Netflix to do it again on ARM after demonstrating the benefits on desktop.

bawolffreply

> It won’t be know until it’s been tried We can reevaluate when and if improvements are made. The problem with statements like this, is its really easy to miss the showstoppers when imagining what something could be. All plans are amazing until you have to actually do them in the real world. As the saying goes: Never fall in love with potential.

izacusreply

"We don't want to support another pile of security bug ridden C++ library" absolutely aren't "dubious grounds". You'd rip a new one to Google if there would be a CVE in a new C++ library in Chrome because of it. Now that Rust library is available, they will continue adoption, as it should be.

account42reply

> "We don't want to support another pile of security bug ridden C++ library" absolutely aren't "dubious grounds". It is when that didn't stop them YOLO'ing in webp and then avif support.

Daiz

A potentially major issue I have with AVIF is that because it is based on a video format, any hardware decoding support AVIF will get is likely to be restrained to common video scenarios. This can result in eg. only 4:2:0 YUV being supported by hardware decoders, as that's the upper limit of AV1 Main Profile (and thus the limit of AVIF Baseline Profile). And 4:2:0 is just a poor fit for certain categories of images (like various kinds of illustrations, screenshots, etc). I would not be very happy about a web where lossy 4:4:4 image compression would be heavily discouraged with modern formats. Also, good lossless compression absolutely does matter for the web as well. Lossy images just cannot be used in all circumstances - like when comparing compression quality (especially of videos), for example! EDIT: To give an additional example, pixel art is an entire field of art (which is very much still active today) where both lossless compression is a must and where 4:2:0 would be absolutely catastrophic for quality.

torginusreply

I'm not super familiar with AV1/AVIF but I do have extensive experience using h264 for desktop streaming, and all major implementations support 444 in hardware and software, and I doubt more advanced codecs/image formats have a limitation like this. Asked ChatGPT and it also claimed there's no problem with 444 and AVIF/AV1.

lgkrnkglwnreply

Then chatgpt missed an obvious issue, nvidia definitely have had shortcomings in the hardware implementation of coder/decoders only supporting 420 in nvenc/nvdec. One obvious problem with that was that it affected premiere.

torginusreply

We have been using NVENC with 444 and h264 in production going back years and HW as old as Turing. But it turns out you are right, 444 decoding with H264 doesn't seem to be supported up until recently: https://developer.nvidia.com/video-encode-decode-support-mat... Never noticed, seems like a strange omission, but in any case doesn't seem to apply to newer codecs like AV1.

jaffathecakereply

I'm pretty sure 444 h264 fails to decode on iOS Safari due to poor support in the hardware decoder.

MrSqueezlesreply

AVIF's support for non-photographic images is covered in the article. It supports lossless.

Daizreply

Yes, I know the format supports it. The question is, will hardware decoders built primarily around video use cases? And if those kind of hardware decoders end up forming the majority of hardware decoders for AVIF, will we end up in a future where using the full range of the format is heavily discouraged because it won't be supported by hardware decoders? For comparison, it was possible to do 10-bit video with H.264, but 10-bit video didn't see any mainstream adoption with it because hardware decoders generally didn't support 10-bit H.264 video. Only when newer formats came around and specced 10-bit support as a baseline necessity did we start to see wider adoption for it. And well, the baseline specs (main profile) for AV1, which AVIF is based on, are limited to 4:2:0...

edflsafoiewqreply

AFAIK WebP has never used hardware decoding and no browser uses it for AVIF now.

account42reply

It's not lossless if you have to convert to YUV first.

jaffathecakereply

You are correct. However, AVIF does not require this conversion. For lossless it supports an "identity" matrix coefficient, which means no conversion. It isn't particularly efficient in terms of file size, but as others have said, lossless images within a web page is extremely niche.

miladyincontrolreply

Wait, DAIZ!? I genuinely did not expect to see you here chiming in. Still appreciate much of your 10 bit anime advocacy from way back.

aniviacat

> Progressive rendering (which AVIF supports) decodes a low-fidelity rendition before the full image arrives. AVIF didn't support progressive rendering for a while, and during that time I believe it was deeply oversold. Now that libavif has implemented it (it was always possible), the conversation appears to be over. This is false. AVIF does not support progressive decoding/rendering. What the author is referring to here is image layering. A lower resolution image layered below the full resolution image, which is loaded and rendered first. That is not progressive decoding/rendering. It is a thumbnail. Unlike AVIF's false advertising, JXL truly supports progressive decoding/rendering. With JXL, you do not first load a thumbnail before loading the actual image. The lower-quality image that JXL shows while decoding is derived from the data of the actual, full-resolution image.

computerbusterreply

Not sure if the difference is materially relevant to UX at all. JPEG XL achieves progressive rendering at a great cost to its selection of coding tools, so I side with AVIF's approach.

MiroslavPokornyreply

How does progressive rendering actually that most OS GUIS are not multithreaded ? The preview in finder is still going to block. Im saying this as a MAC user speaking of experience browsing mounted drives that sometimes are slow and it feels like the window freezes up.

wmfreply

Progressive rendering is intended for the Web and it does work.

brigadereply

Thumbnailing happens in a background process and shouldn’t block the main thread; rather hangs are usually because fetching xattr over SMB is slow and Finder will block until it’s enumerated the metadata of every file in a folder. Preview actually does use progressive rendering for large enough images, even from SSDs. You can tell because opening an image will be blurry for several seconds, even when the full size image could have been decoded in a quarter second or less.

brigadereply

You're wrong. AVIF does support spatial layering where the full-resolution image is derived from the lower-resolution layer, and cannot be decoded independently. Yes, the format is not inherently progressive. But having the fine steps in progressive quality that JPEG-2000 and JPEG-XL have has less general usefulness than the amount of words complaining about it.

F3nd0reply

Can a similar technique be employed in JPEG XL, or does the format not allow for it in any way?

lgkrnkglwnreply

Jpeg2000 (and jpeg ls and jpeg xr) thankfully wasn’t implemented in any major way. Keep it simple, stupid.

nneonneo

That JPEG XL prime computation is a pretty ugly DoS. Just selecting it in Finder, with the preview pane open, maxed out every core on my Mac inside a QuickLookSatellite that also ate 4GB of RAM while doing so - for a good 15 seconds. Not bad for a 2KB picture. It seems like Apple did not set sane limits on their JXL previewer. It is, however, an incredibly cool demo of what the format is capable of. I'm not completely sure if an image format should be that flexible, but I'm impressed nonetheless.

estreply

I wonder if similar hacks apply to zlib and .png as well.

nneonneoreply

Probably not; for PNG, the image size is declared in the header, so a decoder can decide immediately if it wants to decode the image or not. The output is bounded by the size of the image times the bit depth, and decompression runs in time proportional to output size. zlib bombs exist, but they don't affect png because a decoder can simply refuse to decompress past the size of the pixel buffer.

tacomagickreply

The infamous PIL DOS errors, rooting from the library refusing to process images larger than a hard limit. Does this mean though JXL does not have that data accesible quickly?

wmfreply

Zip and PNG bombs have been around for a while: https://github.com/0x48piraj/gz-bomb https://libpng.sourceforge.io/decompression_bombs.html

estreply

I mean "mild" bombs where the final output bytes looks normal, a regular sized image, but the decompressing steps takes unnecessarily long, eating CPU or/and RAM

erureply

You can also have a zip that decompresses to itself: https://research.swtch.com/zip

dylan604reply

When a format has so much flexibility, the world will glom onto the first thing that it solves and put it into the wild to solve that problem. The rest of the capabilities fall by the way side, yet not removed from the format. They're just ignored. MP4 can do so much more than the typical deliverable of a video stream and an audio stream. The spec allows for multiple video streams, multiple audio streams, subtitles, Flash like interactivity to allow self contained DVD style programming of menus to allow for chapter navigation, audio/sub selection, multiangle, etc.

fc417fc802reply

> MP4 can do so much more None of that has fallen by the wayside though? I have seen examples of all of those in the wild except for interactive menus and multiangle. Multiple video streams is incredibly rare to encounter but I have run into it a few times. The argument in favor of flexibility (and jxl) is that if you optimize things for the "average" web user (as the essay seems to be suggesting) then fairly mundane usecases require you to start juggling formats, support becomes spotty, and things start breaking. It's nice to have generous limits within which you can be confident that things will "just work" for the end user. Even just on my own system I'd much rather use a single format rather than dealing with app x not supporting format y. tl;dr jxl is the mp4 of image formats and that's exactly why I like it. The only things I agree with the essay about are progressive decoding and decoding speed. Particularly the latter badly needs to be improved.

v2026.09.14

Weather

World Clocks

Markets

v2026.09.14

Weather

World Clocks

Markets

The Debate

Is human social order fundamentally constructed through shared symbolic fictions, or is it an emergent reaction to physical and ecological coordination constraints?

Should Congress enact a statutory moratorium on training AI models above a specified compute threshold until mandatory safety protocols are established?

Is personal identity over time constituted by psychological continuity, or does it require biological persistence of the same physical organism?

From the Labs

Larry Ellison Cancels $7.5 Billion Planned Sale of Oracle Stock

Google Research Introduces ToolGrad for Synthetic Tool-Use Data Generation

OpenAI Delays IPO Plans Past 2026

Hugging Face Connects Async GRPO and LoRA on HF Jobs

Insight Partners Diversifies Venture Strategy Beyond Frontier AI Labs

Y Combinator Demo Day Highlights Deep Tech Startups Across AI and Hardware

AI Safety Resignations and Doom Warnings Prompt Industry Debate

Obama Urges House Democrats to Build Clear AI Governance Framework

Lyft Launches Commercial Driverless Robotaxi Service with Waymo in Nashville

Fusion Power Startups Partner with Defense Contractors and Military Agencies

Matt Mullenweg Retains Automattic CEO Role Following Board Ouster Attempt

Anthropic CEO Outlines Proposal to Slow Frontier AI Development

Tesla sets October 1 launch date for second-generation Roadster

Hacker News

Ask HN: What are you working on? (September 2026)

Fable 5.1 Solves the Cyphral Distich, a 370-year-old cipher

Why is Google still serving dodgy ads?

Flock worker calls police on reporter filming public camera installation

David Sacks: OpenAI and Anthropic Don't Need Regulations to Pace Frontier Models

The case against JPEG XL

The Rabbit Hole

2026 Swedish general election

Research

Generative Marketing Mix Modeling: A Causal Inference Framework Linking GEO and GEM to Business Impact

Buyer Artificial Intelligence-Enabled Environmental Governance and Supplier Environmental Controversies: An Organizational Information Processing and Signaling

Hierarchical Clustering Can Jointly Satisfy Richness, Consistency, and Scale Invariance

From the Bookshelf

The Art of Doing Science and Engineering

Hamming's Question

Contrarian Corner

Standard AI alignment techniques might make models more dangerous by hiding bad behavior rather than fixing it.

Smartphones and screen time are likely not the main cause of recent student test score declines.

Scheming AI models can still be caught cheating by probing their internal neural activations.

Photograph

On This Day

Games

Puzzles

The Four Rooms

The Weighing of the Stones

The Shifted Sentence

The Track

Ayer

Quote

Word

Komorebi

Albedo