v2026.08.30

Sunday, August 30, 2026

partial

The Marías

Cariño

0:004:36

Weather

Bloomington

66°

Clear

H 89°L 65°

Feels like: 68°
Humidity: 86%
Cloud: 0%
Rain: 2%

Today

89°

65°

Mon

92°

71°

Tue

99°

72°

Wed

100°

74°

Thu

100°

72°

Fri

99°

72°

Sat

88°

69°

World Clocks

San Francisco

—

New York

—

London

—

Berlin

—

Dubai

—

Mumbai

—

Shanghai

—

Sydney

—

Markets

Software and big tech platforms led the day, with ServiceNow surging 4.54% and Amazon gaining 3.97%, alongside modest rises in Microsoft and Alphabet. Semiconductor stocks anchored the bottom as Nvidia dropped 4.57%, pulling down Intel, AMD, and TSMC alongside a slump in crypto assets like Bitcoin. The session highlighted a sharp division across tech, as investors shifted away from chipmakers and hardware names into software and cloud services.

Version v2026.08.30 - Daily Briefing

The Debate

Did Neandertals independently develop symbolic material culture before the arrival of Homo sapiens in Europe?

Settling whether symbolic cognition arose uniquely in modern humans redefines our understanding of human uniqueness and hominin evolutionary trajectories.

Independent Innovation

Uranium-thorium dating of carbonate crusts over cave paintings in Spain, such as at Maltravieso and Ardales, yields dates around 64,000 years ago—at least 20,000 years before modern humans arrived in western Europe. Coupled with marine shell beads colored with ochre found at Cueva de los Aviones, the timeline indicates Neandertals possessed full cognitive capacity for symbolic representation. They harvested specific bird feathers and raptor talons for ornamentation millions of years after splitting from our common ancestor. This evidence suggests that abstract thought, semiotics, and symbolic encoding were not exclusive gifts of Homo sapiens, but rather shared ancestral hominin traits or parallel cognitive developments triggered by complex social organization.

Accidental or Contaminated

Critics challenge the uranium-thorium dating method applied to cave walls, noting that open-system leaching of uranium can artificially inflate calcite crust ages, making 35,000-year-old pigment layers look 65,000 years old. Furthermore, across vast swathes of Eurasia inhabited by Neandertals for over 300,000 years, symbolic artifacts remain astonishingly rare and spatially isolated compared to the explosive, ubiquitous flourish of figurative art, musical instruments, and personal adornment that accompanied Homo sapiens everywhere they migrated. The vast majority of Neandertal toolkits remain strictly functional and static over millennia. What appears to be pre-contact symbolic behavior may instead represent sample contamination, natural mineral staining, or taphonomic anomalies misconstrued as intentional cultural production.

cruxDefinitive micro-stratigraphic dating of untouched pigments combined with widespread, indisputable pre-45,000-year-old symbolic sites across multiple Neandertal regions would settle the debate.

Does the predictive success of idealized scientific models justify believing that their unobservable entities actually exist?

Our answer dictates whether science uncovers fundamental truth about unobservable reality or simply builds increasingly refined tools for human navigation.

Scientific Realism

The miraculous success argument points out that if theoretical entities like quarks, gravitational fields, and electron orbitals were mere fictions, the precise predictive accuracy of modern physics would be a cosmic coincidence. Models use idealizations—like frictionless planes or infinite populations—to strip away noise, isolating the true causal structures that drive real-world phenomena. When independent lines of experimental intervention consistently manipulate these unobservable entities to produce intended macroscopic effects, rejecting their physical existence requires an implausible skepticism. Success in predicting novel phenomena before empirical observation proves that our best models successfully latch onto the hidden architecture of reality, even if our descriptions remain approximations undergoing continuous refinement.

Instrumentalism

The history of science is a graveyard of wildly successful theories that posited entities later shown not to exist, from caloric fluid to the luminiferous aether. Idealized models deliberately introduce known falsehoods—treating continuous fluids as smooth continuums or assuming point masses—meaning the model itself is false by design. Different, mutually incompatible models often predict the exact same empirical observations with equal precision, a problem known as underdetermination. Because we only ever measure macro-level inputs and outputs, inferring the literal existence of invisible entities relies on an unwarranted jump from predictive utility to ontological truth. Models function as intellectual scaffolding rather than mirrors of hidden reality.

cruxA rigorous proof showing that no two structurally incompatible models can generate identical empirical predictions across all possible observation regimes would settle the conflict.

Can transgenerational epigenetic inheritance drive permanent adaptive evolution in complex animals without changing DNA sequences?

If inherited epigenetic marks permanently alter evolutionary trajectories, the standard Modern Synthesis must expand to include non-genetic inheritance as a core driver of adaptation.

Epigenetic Driver

Small RNA molecules, histone modifications, and DNA methylation patterns can pass down through germ cells across generations in response to environmental stress, altering phenotypes without changing underlying nucleotide sequences. In organisms like Caenorhabditis elegans, stress-induced small RNAs persist for dozens of generations, providing immediate adaptive flexibility faster than random DNA mutations can occur. By steering phenotypic variation in specific directions, persistent epigenetic changes buy populations critical time in harsh environments, effectively guiding where natural selection acts. If these marks lock in phenotypic shifts long enough for genetic assimilation to follow, epigenetic inheritance acts as an active, directed engine of evolutionary change rather than a temporary physiological echo.

Genetic Primacy

Transgenerational epigenetic marks are overwhelmingly erased during germline reprogramming in early embryonic development, resetting the genome for the next generation. The rare epigenetic marks that survive past two or three generations in laboratory models like worms rapidly decay, making them far too unstable to sustain evolutionary adaptations over deep time. Natural selection requires stable traits preserved across hundreds of generations, a fidelity that only double-stranded DNA sequence changes provide. What looks like adaptive epigenetic inheritance is often either direct environmental exposure of germ cells inside the parent or background genetic variation masking as non-genetic inheritance. DNA remains the sole high-fidelity medium of macroevolution.

cruxIdentifying a stable, non-genetic chromatin modification in wild animal populations that persists for hundreds of generations without DNA sequence changes would settle the issue.

From the Labs

Sony Music Publishing/lab

Sony Music and Warner Chappell Sue Anthropic Over Copyrighted Works

Late Friday in the U.S. District Court for the Northern District of California, Sony Music Publishing, Warner Chappell, and other publishers filed a lawsuit against Anthropic and co-founders Dario Amodei and Benjamin Mann. The plaintiffs allege a campaign of illegally torrenting, scraping, and downloading thousands of copyrighted works, including sheet music and lyrics, to train Claude. This builds on prior litigation like the Bartz v. Anthropic case, where Anthropic was ordered to pay 1.5 billion dollars after a judge ruled that while using copyrighted material for training was legal, acquiring it via piracy was not.

This lawsuit marks a sharp shift in AI litigation by focusing explicitly on the exact acquisition methods of training data rather than just the final model output. If the courts maintain that fair use covers training but draw a hard line at torrenting or unauthorized downloads, AI labs will face massive financial and operational penalties for how they gather data. Anthropic says it disagrees with the claims and plans to defend itself in court, but with lawyers from Concord Music Group and Universal Music Group involved, the legal pressure is mounting. The ultimate boundary between legal scraping and actionable piracy remains the core battleground for future generative models.

TechCrunch

Hugging Face/lab

Hugging Face Adds Hindi and Indian English Sets to Open ASR Leaderboard

Voice Arena and Hugging Face partnered to launch Monsoon en-IN and Monsoon hi-IN, adding the first Indic languages to the Open ASR Leaderboard. The dataset contains unscripted dual-channel spontaneous conversations split into speaker-disjoint public and private sets across 4,888 speakers, with no single speaker accounting for more than 6.8 percent of total duration. The collection covers 12 recorded metadata attributes per clip including age, gender, occupation, income band, handset model, and district. Because standard text normalizers fail to map Hindi spelling variations cleanly, the Hindi dataset ships a lattice of accepted spellings for each transcript span rather than relying on a single string reference.

Traditional speech recognition benchmarks rely on limited hardware and homogeneous audio sets, masking severe performance disparities across demographics and accents. By drawing Indian English from 428 native districts across 30 states and territories and sampling Hindi across 202 to 295 districts, Monsoon prevents models from overfitting to specific hardware or regional accents. Automatic Speech Recognition, or ASR, is the underlying technology that converts spoken audio into written text. Evaluating models against diverse real-world voices ensures that speech recognition systems work reliably for over half a billion Hindi speakers instead of just small subsets. Whether developer adoption will follow these new evaluation metrics remains to be seen.

Hugging Face

Google Research/lab

Google Research introduces planetary prediction engine for global models

Google Research engineers Rama Pasumarthi and Shravya Shetty introduced the planetary prediction engine, an autonomous AI system within Google Earth AI. The system automates end-to-end geospatial modeling from natural language queries across data discovery, feature engineering, model training, and reporting. To bypass context-window limitations, it uses three modular stages orchestrated by off-the-shelf LLMs that exchange opaque data handles instead of raw text. Across 21 CDC health indicators, the system hit a mean R² of 76.8% versus 60.0% for manual expert pipelines. In testing during the 2026 Bundibugyo ebolavirus outbreak in the Democratic Republic of the Congo, it achieved a Recall@10 of 83.3%, correctly identifying 15 of 18 new transmission hotspots across five weekly forecasts.

This automated approach cuts geospatial modeling times from weeks of manual data engineering down to minutes. In crisis response, rapid predictions are critical; during disease outbreaks or natural disasters, delay costs lives. The system also solves data resolution problems in resource-scarce regions. Downscaling food security data in Nigeria from provincial ADM1 levels to local ADM2 levels doubled baseline prediction accuracy to an R² of 66.1%. The results prove the strength of pairing structured statistical covariates with latent foundation model embeddings like AlphaEarth. Still, because this remains an experimental research capability, real-world deployment efficacy across unpredictable global crises remains to be fully verified.

Google Research

TechCrunch/startup

Hollywood celebrities enter microdrama market through apps and investments

Hollywood actors and producers are moving into microdramas, short vertical series built for smartphone screens that originated in China and are projected to hit $1.5 billion in US revenue this year. James Franco stars in Shortical's crime-and-werewolf series Love, Lies & Frank, produced under SAG-AFTRA's new low-budget verticals agreement. Issa Rae's Hoorae Media partnered with TikTok's PineDrama app to release Screen Time, a thriller that crossed 150 million views after hitting 75 million in its first week. Meanwhile, Taye Diggs starred in CandyJar's Off Limits & All Mine and launched his own vertical platform, Microhouse Films. Additionally, Kim Kardashian and Kris Jenner invested in GammaTime.

Traditional movie studios are seeking cheaper, faster production methods while actors face employment shortages and short-form platforms like TikTok and YouTube Shorts consume audience attention. Microdramas offer brief, high-drama episodes designed for quick monetized viewing, opening new income streams under specialized union contracts. The shift is broadening beyond standard soap opera formulas; for instance, Jesse Tyler Ferguson is developing a zombie musical comedy titled Theater Kids Vs. Zombies for vertical platform aTwist. However, it remains uncertain whether premium mobile microdramas can sustain long-term user retention once initial celebrity novelty wears off, or if mainstream audiences will pay consistently for minute-long episodes.

TechCrunch

Google DeepMind/lab

Google DeepMind releases Gemini Omni 1.1 Flash with enhanced video controls

Google DeepMind released Gemini Omni 1.1 Flash, an updated video generation model available through Google AI Studio and the Gemini Enterprise Agent Platform. The model adds scene extension capabilities, allowing developers to extend existing clips in 10-second increments up to a total length of 40 seconds. It can analyze up to 10 seconds of prior visual context, a significant increase from the single second of context in earlier versions. Developers can also set specific starting and ending keyframes to orchestrate camera movements like orbital rotations, whip-pans, and snap-zooms. For faster iteration, the update introduces a 360p preview mode that renders up to 60 percent faster and costs one-third as much as standard 720p generation, alongside crisp 4K output.

Giving creators explicit frame and camera control shifts generative video from a fun gimmick to a practical software component. Being able to lock a character in place while the background stretches or to specify an exact landing frame makes seamless looping and multi-shot storytelling actually work without random jumps. The addition of low-cost 360p draft rendering lowers the financial barrier for testing complex prompts before committing compute to final 4K renders. What remains to be seen is how reliably the model maintains long-term character coherence across the full 40-second limit, as well as how smoothly third-party creative software platforms will integrate these specific API interactions into existing video editing timelines.

Google DeepMind

Therabody/startup

Therabody launches Theragun Sense massage gun with guided LCD routines

Therabody launched the Theragun Sense (2nd Gen), a $299 wellness massage gun designed for everyday aches and mild muscle tension. Unlike heavy-duty deep-tissue devices, this model focuses on lighter relaxation across five adjustable speed settings. Its standout feature is an onboard LCD screen that walks users through four built-in routines, providing visual instructions on placement, duration, and applied pressure. Users can pair the device over Bluetooth with the Therabody app to load additional guided sessions, including a six-minute Sleep routine or targeted Desk Relief. The app also features an AI tool called Coach that offers personalized recovery plans based on daily activities. It ships with USB-C charging, a travel pouch, and two standard attachments.

Perfect execution on basic usability solves the biggest issue with consumer wellness hardware, which is that people buy them and immediately forget how to use them. By putting step-by-step visual timers directly on the handle screen, the device removes the guesswork of target areas and pressure application. It also runs quietly enough to use while watching television without creating annoying background noise. However, buyers looking for serious deep-tissue therapy for severe muscle knots will find the motor power underwhelming compared to heavier devices like the Theragun PRO. It remains uncertain whether casual users will stick with the app-connected AI features long term once the novelty of guided recovery routines wears off.

TechCrunch

Google DeepMind/lab

Google DeepMind pilots double-blind AI evaluations using hardware enclaves

Google DeepMind partnered with the Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons to run a double-blind evaluation on a Gemini Flash Lite model. To prevent benchmark contamination where a model accidentally trains on test questions, the team used Google Cloud's Confidential Space within its Confidential Computing portfolio. This cryptographic setup creates an isolated execution environment where external benchmarks run against proprietary weights. Under this protocol, external evaluators cannot access or view the model weights, while Google cannot see the evaluator's private test prompts. This cryptographic proof ensures both data privacy for evaluators and intellectual property protection for model providers, replacing traditional non-technical zero-logging legal agreements with hardware-level security guarantees.

AI benchmark scores are notoriously easy to fake or inflate when test sets leak into training data, making independent safety claims hard to verify. Moving from contractual promises to cryptographic isolation fixes the structural trade-off between inspecting a model and protecting its underlying weights. This approach opens up safe pathways for government bodies and cybersecurity auditors to evaluate proprietary systems against highly sensitive, classified prompts without risking data leaks. What remains to be proven is whether hardware-enclave evaluations can scale efficiently across massive model sizes without imposing crippling performance overhead, and whether independent safety institutes will universally adopt this specific technical protocol as an industry evaluation standard.

Google DeepMind

XREAL/essay

XREAL a01 smart glasses deliver vivid displays alongside thermal and utility trade-offs

XREAL released the a01, a $300 smart glasses model that functions as a tethered external monitor rather than a standalone augmented reality device. The lightweight glasses lack an internal battery or operating system, requiring a USB-C connection to a host device like a handheld gaming console, phone, or laptop. Visually, the device uses dual mini OLED panels delivering 1080p resolution and up to 1,600 nits of brightness, simulating a 147-inch projected display. Users can also pair it with the $200 Beam Pro mini-tablet accessory for standalone media streaming without relying on a connected phone or PC.

While the a01 produces vibrant imagery, it highlights the practical friction facing wearable displays. Screen-mirroring creates awkward redundancies, such as requiring a phone or laptop screen to stay illuminated inches away while watching content inside the glasses. Furthermore, thermal management remains an unsolved hardware constraint across light XR devices. Compact electronics processing video content on your face generate noticeable heat, producing a warm, uncomfortable tingling sensation across the bridge of the nose and eyes during extended use. Until manufacturers solve heat dissipation and screen independence, these glasses remain niche accessories rather than replacements for traditional screens.

TechCrunch

Google Research/lab

Google Research introduces GlucoFM for continuous glucose monitoring

Google Research released GlucoFM, a self-supervised dual-stream foundation model for continuous glucose monitoring (CGM) data. Pre-trained on 109,066 hours of unlabeled CGM data across 477 records from Wear-CGM and four other datasets, GlucoFM aligns readings to a 24-hour, five-minute grid with an observation mask. Its dual-stream encoder splits low-frequency baseline glycemic trends from short-term residual event deviations. Using latent predictive pre-training and CGM-aware data augmentations like drift and drops, GlucoFM was evaluated across four cohorts—CGMacros, Stanford, Hall, ShanghaiT2DM—and seven clinical tasks. It outperformed the best-evaluated GluFormer variant, scoring a 5.8 percentage point higher average PR-AUC across 14 cohort-task evaluations while achieving the lowest mean absolute error in postprandial glycemic response forecasting across Dexcom and Libre devices.

Standard continuous glucose monitors generate streams of interstitial glucose data, but making sense of them usually requires expensive clinical labels. Most existing foundation models treat glucose traces as a single undifferentiated stream, mixing slow baseline shifts with sudden spikes from meals or activity. By decoupling baseline trends from short-term deviations, GlucoFM extracts cleaner health signals without needing raw reading reconstruction, which is vulnerable to sensor noise. The model demonstrates high cross-dataset transferability and strong few-shot performance when labeled patient data is sparse. What remains uncertain is how well these dual-stream representations generalize to even broader consumer populations outside clinical cohorts, or how effectively this approach integrates with other wearable health metrics like activity and sleep.

Google Research

Lambda/funding

Lambda secures $1B in debt to purchase Nvidia GPUs for Microsoft

AI cloud infrastructure provider Lambda secured $1 billion in private, short-dated debt to purchase Nvidia AI chips, which will be leased directly to Microsoft. Arranged by JP Morgan Chase, the short-dated structure indicates Lambda expects to deploy the hardware rapidly and service the debt using incoming lease revenue. This transaction follows a busy financing period for Lambda, including a $1 billion secured credit facility in May and a $926 million loan closed this week to fund Nvidia GB300 GPUs under contract. The company raised $1.5 billion in venture funding at a $5.43 billion post-money valuation in November, and is currently reported to be in talks for a $3 billion pre-IPO funding round.

Lambda's deal reflects a broader trend where specialty cloud providers rely heavily on debt to secure critical hardware in the ongoing AI expansion. According to Bloomberg data, global AI-related debt issuance by banks and tech companies has topped $400 billion so far in 2026. Taking short-dated debt allows Lambda to bypass traditional equity dilution while matching borrowing costs against immediate revenue from tech giants like Microsoft. However, this capital-intensive model relies on continuous, high-margin demand for GPU compute. If deployment delays occur or customer lease demand shifts faster than expected, servicing massive short-dated obligations could quickly create financial pressure for specialized cloud operators navigating rapid hardware replacement cycles.

TechCrunch

TechBBQ/essay

European Tech Leaders Debate AI Sovereignty at Copenhagen TechBBQ

At the annual TechBBQ conference in Copenhagen, which concluded on August 27, European investors, founders, and executives gathered to discuss technological sovereignty under the theme Emerging from Agency. The urgent focus on owning infrastructure followed an outage where Anthropic's AI models, Mythos and Fable, became unavailable outside Europe earlier in the year. Speaker panels featured Signal President Meredith Whittaker, who criticized OS integrations like ChatGPT in iMessage for creating data collection apparatuses, alongside Stability AI co-founder Emad Mostaque, who warned that concentration of power means whoever controls AI will effectively control countries. Ellen de Brever from the Novo Nordisk Foundation Cellerator also emphasized human judgment in delegating agentic tasks.

Relying on U.S. and Chinese technology leaves European tech startups exposed to sudden service disruptions and foreign regulatory shifts beyond their influence. While some software teams were heavily disrupted by the Anthropic outage, others view geopolitical tech dependency as a manageable risk for now. However, as AI agents take over more cognitive tasks and robotics handles physical work, fundamental questions about economic participation, data privacy, and governance come to the forefront. The push for local models and infrastructure reflects a growing demand for privacy and independence, but European ecosystems still face the steep challenge of raising sufficient growth capital to build true operational alternatives.

TechCrunch

VZVC/funding

Vijay Pande launches VZVC to make concentrated bets in AI biotech

Vijay Pande, former Stanford chemistry professor and creator of Folding@home, left Andreessen Horowitz after growing its healthcare practice to nearly $4 billion over a decade. He co-founded a new firm named VZVC with investor Zach Werner. Instead of making dozens of investments annually, VZVC focuses on a few concentrated bets per year, operates without associates, and relies heavily on AI for daily operations. Pande argues AI shifts drug discovery from serendipity to engineering by improving target identification, drug design, and clinical trials. He notes that traditional drug development relies on animal models like mice that poorly predict human outcomes, causing an 80% failure rate from first to final trial phase.

The shift matters because clinical trials cost hundreds of millions of dollars, and an 80% failure rate drives up the amortized cost of medicine. AI models offer higher predictive accuracy than animal models, promising to lower trial failures and cut overall drug costs. Furthermore, precision medicine is evolving beyond static genomic blueprints to track real-time proteomic changes and automated measurements tailored to individual baselines rather than population averages. However, biological data cannot simply be scraped from the internet like text datasets. Because biotech companies typically build proprietary, walled-off datasets, it remains uncertain how broadly these AI advances will be shared across the medical ecosystem.

TechCrunch

Nvidia/lab

Nvidia shifts focus to data orchestration and whole-system infrastructure

Nvidia is pivoting its edge from standalone GPUs to full-system orchestration with its new Vera Rubin architecture. The system combines the Rubin GPU with the Vera CPU, the Groq 3 LPX inference accelerator, and dedicated storage and networking racks to manage gigawatt-scale data center bottlenecks. According to Jason Hardy, Nvidia’s VP of storage technology, the Vera CPU handles data traffic control to feed GPUs faster without memory bottlenecks, yielding up to a 3x speedup in storage operations. By contrast, OpenAI designed its Jalapeño chip to keep whole workloads within a single connected system to avoid data movement altogether. Both approaches tackle the same physical limit of getting data to processors efficiently.

Compute power inside megascale data centers is becoming a commodity, but running these facilities at peak efficiency as they reach gigawatt scales remains extremely difficult. Driving down tokens-per-watt now depends heavily on traffic direction outside the GPU rather than just raw processor speed. Nvidia’s new ecosystem approach allows it to defend its market dominance even as hyperscalers like Google and Amazon build custom competing GPUs. However, moving the battleground to data orchestration does not guarantee an easy victory. Nvidia must now compete against rival chipmakers and cloud providers across an entire secondary layer of infrastructure, where hardware integration matters more than raw chip performance.

TechCrunch

Xpeng/funding

Chinese automakers enter humanoid robotics with massive private funding rounds

Chinese electric vehicle makers are rushing into humanoid robotics, driven by falling automotive margins and advances in AI. Xpeng’s robotics unit raised over $900 million at a post-money valuation exceeding $6.3 billion, marking the largest single-round private financing in China’s embodied AI sector. The round was led by IDG Capital, alongside Gaorong Ventures, Tencent, Alibaba, and $100 million from Xpeng founders He Xiaopeng and Brian Gu. Xpeng’s commercial humanoid, Iron, joins competing efforts like Chery’s AiMOGA unit preparing for an IPO, BYD’s Xiao Di robot, and Hyundai’s plan to deploy Boston Dynamics’ Atlas in its Georgia factory by 2028. Mobileye also acquired Mentee Robotics for $900 million.

Automakers view humanoid robots as their next major profit engine because car sales face razor-thin margins. The technical thesis relies on using large language model AI techniques to help physical hardware quickly master complex manual tasks. With established supply chains and heavy manufacturing capacity, car companies are uniquely positioned to assemble physical robots at scale. However, physical hardware is only half the battle. While companies like Xpeng have the hardware components to build realistic humanoids like Iron, it remains uncertain whether Chinese automakers can catch up to Tesla’s advanced AI software systems, which dictate how effectively these machines operate in complex environments.

TechCrunch

Anthropic/lab

Anthropic Details Automated AI Researchers for Alignment Post-Training

Anthropic fellow Chen Yueh-Han published a paper titled Automated Researchers Can Reliably Mitigate Alignment Failures, demonstrating an automated system that handles AI training. The Automated Alignment Researcher, or AAR, targets misaligned behaviors across 10 specific benchmarks. It scans available literature, proposes a training method, and runs 30-minute training sessions on the model. The framework keeps effective techniques and discards failures over multiple iterations. Across all 10 alignment benchmarks, the automated system improved performance without degrading the model's overall capabilities. Within six hours on average, the best method generated by the automated system beat out training ideas proposed by experienced human researchers.

This experiment offers practical evidence that automated post-training alignment is closer than expected, moving us toward recursive self-improvement. The economics are stark: running an AAR costs roughly $4 per hour in API inference, whereas Anthropic pays human researchers $150 per hour. If automated agents can refine alignment methods, they could eventually take over broader AI training tasks. However, real bottlenecks remain. The system only succeeds if the underlying benchmarks accurately reflect true alignment goals. Human researchers are still needed to build and update those benchmark tests, as well as curate the scientific literature that the automated system relies on for its proposals.

TechCrunch

Brave/startup

Brave Browser Adds Native Email Aliases for Privacy Protection

Brave launched a built-in email alias feature designed to keep personal email addresses private during online sign-ups. Users register a primary email address with a Brave account, and when clicking any web form email field, a pop-up offers an alias option. Alternatively, users can right-click the field to generate one. Formatted alias emails auto-fill into forms, and incoming messages pass through Brave's servers for basic spam and virus scanning before forwarding to the user's main inbox. Brave deletes the messages from its servers immediately after forwarding. Free accounts receive five email aliases, with additional aliases reserved for future Premium users. Notes accompanying aliases are saved locally unless end-to-end encrypted sync is enabled.

This feature addresses a major tracking vector where ad tech firms like Meta match email addresses across websites to monitor purchases and serve target ads. Aliases also prevent primary emails from being exposed when databases leak to data brokers during site breaches. However, an immediate hurdle exists. Because Brave is new as an email forwarding sender, forwarded messages may initially get filtered into user spam folders until the company builds a stronger provider reputation score. Still, integrating forwarding directly into the browser makes alias creation frictionless compared to third-party services, helping users lock down their digital footprint seamlessly.

TechCrunch

Google DeepMind/lab

Google DeepMind launches Gemini 3.5 Transcribe speech-to-text model

Google DeepMind unveiled Gemini 3.5 Transcribe, a speech-to-text model built for real-time voice applications. The model converts raw audio into formatted text while stripping disfluencies, filler words, and handling background noise and custom vocabulary. Compared to its predecessor Chirp 3, Gemini 3.5 Transcribe cuts time to final transcription by 70% based on Artificial Analysis evaluations. On the multilingual FLEURS benchmark across top languages, it achieved a 5.50% word error rate in streaming mode and 5.04% in non-streaming mode. The model supports multi-speaker attribution and word-level timestamps, and is accessible via the Gemini Live API for real-time streaming and the standard Gemini API for non-streaming workflows across platforms like Google AI Studio.

Speech recognition has long struggled with real-world messiness, forcing developers to build complex post-processing pipelines to clean up filler words or assign speakers accurately. By integrating contextual understanding directly into transcription, Gemini 3.5 Transcribe cleans up speech natively, simplifying integration for developers building voice agents and real-time captioning tools. Google is deploying the tech across its consumer stack, including Android's Rambler feature, Gboard, Antigravity, and Chrome, while developer frameworks like LiveKit and Vercel are adopting the API. However, it remains to be seen how reliably the model maintains low latency and low error rates under extreme background noise or highly domain-specific jargon in complex enterprise environments.

Google DeepMind

Hacker News

486 points451 comments

Debian votes to allow "responsible use of generative AI"

Debian project members voted to clarify their stance on generative AI, establishing that the open-source Linux distribution neither explicitly endorses nor bans AI tools for software development, packaging, or documentation. The policy acknowledges that AI can boost contributor productivity, allowing volunteer developers to focus on higher-level tasks like code review, architectural design, and community collaboration. Crucially, the outcome leaves standard contribution expectations unchanged. Submissions generated with AI assistance must satisfy the exact same benchmarks for quality, maintainability, correctness, and legal compliance as human-written code. The core idea is that tool usage does not dilute personal accountability.

The decision puts the burden of verification entirely on the human contributor, requiring submitters to understand, test, and edit any AI-generated output before submitting it. During the voting process, extreme proposals on both ends of the spectrum failed to gain traction, with several rigid options failing to beat the None of the Above threshold. Debian's pragmatic compromise reflects a practical reality in open-source maintenance, where enforcing complete bans on generated code is nearly impossible. Instead, Debian focuses on output quality over process, ensuring that contributors remain fully responsible for fixing bugs, untangling complex logic, or explaining their code when issues arise down the line.

The community discussion highlights deep anxiety over whether volunteer maintainers will be overwhelmed by low-effort, AI-generated pull requests. Commenters debate whether self-assessed AI labeling systems work or if lazy contributors will simply obscure tool usage, while pointing out that open-weight local models like Qwen3.8 27B offer a privacy-conscious middle ground for developers reluctant to rely on proprietary cloud services.

articlefull thread

From the thread

GZGavinZhao

Related: I find this self-assessed AI level for contributions [1] to be extremely useful, both professionally and personally, to communicate the level of AI assistance so that the person on the receiving end can evaluate how much time & effort they want to spend understanding my code. [1]: https://www.visidata.org/blog/2026/ai/#self-assessed-ai-leve...

chrisjjreply

> this self-assessed AI level for contributions [1] to be extremely useful Why would you trust the bot to self-assess truthfully? That's like trusting a spammer to tag his emails as spam.

zuzululureply

a spammer is motivated by self-enrichment AI agents are grounded on the user provided direction. If it doesn't work well then it naturally loses trust.

chrisjjreply

> If it doesn't work well then it naturally loses trust Too late. If it doesn't work well then it has already naturally destroyed trust.

Calazonreply

This is great. There's a big difference between levels 2-4 (where I generally operate) and levels 7-8 (where I see a lot of others operate), and I like how this level system makes that explicit.

stillpointlabreply

330 points438 comments

Iceland votes on whether to restart talks on joining EU

Icelanders are voting in a tight referendum on whether to restart European Union accession negotiations, thirteen years after talks were suspended in 2013. Prime Minister Kristrún Frostadóttir brought the vote forward amid escalating geopolitical tension, though domestic debates focus heavily on national sovereignty and marine resources. Fishing accounts for nearly 40 percent of Iceland's exports, making the EU Common Fisheries Policy a central sticking point. Opponents fear surrendering control over local waters, remembering past conflicts like the Cod Wars with Britain, while proponents point out that Brussels has signaled potential exemptions to protect Iceland's strictly managed, highly profitable fishing grounds.

Although Iceland already belongs to Schengen and the European Economic Area single market, full EU membership would eventually require adopting the euro currency and joining the customs union. Pro-EU advocates argue that adopting the euro could stabilize Iceland's economy, where inflation has reached 5.6 percent and interest rates sit at 8 percent. However, a referendum victory only authorizes opening negotiations, which would still require completing 35 separate policy chapters, a second public vote, constitutional amendments, and unanimous approval from all 27 EU member states. Security concerns also loom, as the nation maintains no standing army and relies on NATO allies.

Commenters analyze the practical mechanics of Icelandic voting and EU accession, noting how low-tech paper ballot counting provides transparency and public trust. The discussion focuses on why fisheries are an existential economic issue for Iceland rather than mere protectionism, alongside debate over whether non-member associate status or full EU integration offers the best protection against currency instability.

article

425 points289 comments

The internet is kind of a predatory cesspit now

Stephen Diehl looks back at the 1990s web as a chaotic, unpolished public square assembled by obsessive amateurs. It certainly had its share of early flaws, like Usenet flamewars, ugly site tags, viruses, and crude scams, but it was stubbornly human and mostly lacked a financial business model. Today, predation is no longer an occasional abuse of the network; it has become its central organizing principle. Modern platforms continually profile human vulnerabilities, optimize persuasive pitches, and place payment links right next to everyday anxieties. The modern internet functions as an efficient machine designed to detect emotional weakness, amplify it, and monetize whatever niche insecurity pops up.

The bleak reality is that ordinary consumers have been recruited into the lower ranks of the pyramid. People spend their days promoting investments, supplements, or courses they barely understand, acting as unpaid distributors. The consumer, marketer, and product have merged into a single exhausted individual running on feeds designed around outrage and envy. Diehl admits he was wrong to assume grifters were just a small, rotten minority. Instead, fleecing others has become a participatory career choice. Affiliate marketers sell courses on affiliate marketing to each other, and life coaches coach aspiring life coaches, sustaining the psychological trap because admitting the product is worthless means admitting wasted years.

Commenters debate whether Diehl's bleak assessment reflects broader societal decay beyond tech or merely online cynicism. A prominent subthread calls out the article's own writing, pointing to paragraphs that read like automated LLM output and questioning the author's authenticity.

article

296 points189 comments

Hy4 preview

Tencent released Hy4 preview, an open-weights large language model featuring 770 billion total parameters, 49 billion active parameters, and a context window exceeding one million tokens. Designed for productivity tasks in coding, office work, and scientific research, the model scored 2.99 out of 4.00 in Tencent's internal expert benchmark, edging past GLM-5.3 and Kimi K3. It is accessible globally via APIs on Tencent Cloud TokenHub and OpenRouter, as well as integrated into tools like WorkBuddy and CodeBuddy. API pricing is positioned competitively at $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million tokens for cache hits.

The release stands out because Hy4 actively participated in its own development pipeline. It autonomously proposed training methods, data strategies, evaluation frameworks, and low-level operator optimizations, establishing a recursive self-improvement loop during its creation. Furthermore, the model analyzed inference bottlenecks, applying operator fusion and communication tweaks that boosted end-to-end throughput by 31.8% over baseline systems across varying context lengths. Hy4 can also generate playable game prototypes from single natural-language prompts and assist with domain-specific research problems across molecular dynamics, condensed-matter physics, and fundamental mathematics.

The thread highlights rapid initial adoption driven by low cache reading costs on OpenRouter while calling out misleading chart formatting in the announcement. Commenters also challenge calling open-weights models open source and point out how self-improvement claims align with recent AI timeline predictions.

article

299 points169 comments

Lawmakers added $1 to car insurance policies. That money paid for Flock cameras

In 2023, Texas passed a bill adding a single dollar fee to every driver's auto insurance policy. The explicit purpose of the fee hike was to fund efforts to crack down on a massive spike in catalytic converter thefts. Three years later, an investigation by The Texas Tribune revealed how an obscure state agency called the Motor Vehicle Crime Prevention Authority actually spent the cash. Rather than funding localized sting operations, the authority directed at least $30 million from those dollar surcharges into building out a massive, automated license plate reader network supplied by private surveillance company Flock Safety.

The state board, whose members are mostly appointed by Governor Greg Abbott, turned that modest fee into a grid of at least 3,200 Flock cameras stretching across Texas from El Paso all the way to the Louisiana border. While state officials claim the system helped clear roughly 1,660 theft cases, the initiative sparked intense backlash regarding privacy rights and state overreach. The program reached a weird breaking point when Governor Abbott issued an executive order blocking state agencies from spending further public funds on Flock cameras, leaving the long-term future of the expansive statewide tracking network in limbo.

The discussion focuses heavily on whether license plate readers actually address catalytic converter theft, with commenters noting that commodity metal prices drove national theft trends far more than camera networks. Others debated circumventing surveillance through corporate vehicle ownership, while pointing out that private camera vendors likely rely on broader facial recognition and data sharing anyway.

article

154 points165 comments

Highest-ever ocean temperature measured as powerful El Niño forms

Data from Europe's Earth observation agency Copernicus shows that the global average sea surface temperature hit 70 degrees Fahrenheit on August 22, 2026. That marks the highest ocean temperature ever recorded since daily satellite tracking began back in 1979, topping the previous record set in March 2024. While long-term human-driven climate change provides the baseline warming trend, scientists note that a powerful El Nino event in the tropical Pacific is currently supercharging temperatures. The timing is particularly alarming to climate researchers because global ocean waters typically reach peak heat during March and April following the Southern Hemisphere's summer.

Surpassing peak heat thresholds late in August signals dangerous momentum for global weather systems. Warmer water transfers significantly more heat and moisture into the atmosphere, which amplifies land heatwaves and powers more destructive tropical storms. Additionally, hot oceans lose their structural capacity to absorb carbon dioxide, effectively weakening one of Earth's critical natural carbon sinks. Extreme temperature spikes have already been documented in localized hotspots like the western Mediterranean Sea and near both polar regions. With El Nino continuing to push heat into the oceans, scientists warn that risks to marine ecosystems and coastal infrastructure will accelerate.

Commenters added firsthand observations of shifting marine environments, such as divers noting dramatically warmer waters and unusual species movement in Monterey and the Galapagos. The conversation also wrestled with systemic coordination failures, comparing successful historical efforts like stopping ozone depletion against the complex economic incentives driving fossil fuel dependence today.

article

Research

Economics

Sophistication in GenAI Use: Field Evidence from a Large Firm

Nicholas J. Hallman, Zachary T. Kowaleski, Anu Puvvada et al.

Researchers analyzed over 700,000 artificial intelligence prompts written by roughly 4,000 office workers at a large company over eight months. They wanted to see who actually uses AI well. Senior staff wrote much more sophisticated prompts than junior workers, likely because domain experience helps people ask better questions. Teams handling company-wide strategy were also far better at it than other departments. Surprisingly, formal AI training classes did not make employees any better at prompting, and workers did not naturally improve over time.

Generative AI rewards existing domain expertise, meaning senior staff get far more out of the technology than junior workers.

Social Network Structure, Wealth, and Wealth Inequality Across Cultures

Eleanor A. Power, Monique Borgerhoff Mulder, Samuel Bowles et al.

Researchers studied 3,500 households across 46 distinct societies around the world to see how social connections relate to material wealth. Instead of relying on social media data, they tracked real-world interactions like sharing food, borrowing money, and working together. In almost every community, wealthier households had wider networks of mutual support. They helped more neighbors and received help from more neighbors in return. A household's financial standing was also closely tied to the relative wealth of the people in their social network.

Wealthier households globally tend to build broader networks of mutual aid rather than isolated financial cushions.

AI worsens climate change, integrated assessment shows

Huiying Ye, Richard S. J. Tol, Fangzhi Wang

Economists modeled how the rapid expansion of artificial intelligence will impact global climate change. By integrating AI growth into established climate models, they found that AI development is a net polluter. Depending on whether AI acts like traditional IT software or drives a major industrial revolution, it could add between 0.1 and 0.8 degrees Celsius to global warming by the year 2100. Those added environmental damages would effectively swallow roughly one-fifth to one-quarter of the total economic gains AI creates.

AI's economic gains come with a massive hidden cost, as climate damage could erase a quarter of its profits by 2100.

Puzzles

logicmedium

The Three Storage Units

You have three storage units labeled Apples, Oranges, and Mixed. Each label is incorrect. You are allowed to pull one piece of fruit from one unit to see what is inside. Which unit do you pick from to identify the contents of all three?

Hint →

Since you know every label is definitely wrong, picking from the unit labeled Mixed is your best starting point.

Answer →

Pick from the unit labeled Mixed.

Because all labels are wrong, the unit labeled Mixed must contain either only Apples or only Oranges. If you pull an Apple, you know that unit is actually the Apple unit. Since the label Oranges cannot be Oranges and you just found the Apple unit, the unit labeled Oranges must be Mixed, and the unit labeled Apples must be Oranges. This uses the logic of constrained sets to solve the whole system with one observation.

matheasy

The Salary Hike

A worker earns 100 dollars a day. Their salary is increased by 20 percent on Monday. On Tuesday, the new salary is decreased by 20 percent. What is the final daily salary?

Hint →

Calculate the Monday increase first, then apply the Tuesday decrease to that new total.

Answer →

96 dollars.

A 20 percent increase on 100 makes the salary 120. A 20 percent decrease on 120 is 24, because 0.20 times 120 equals 24. Subtracting 24 from 120 leaves 96. Percentages change value depending on the base number they are applied to.

weekendhard

The Architect's Walk

An architect is designing a courtyard with a square perimeter of 160 meters. Part 1: What is the area of the square? Part 2: She wants to place a fountain in the exact center of the square. If the fountain is a circle with a diameter of 10 meters, what is the area of the courtyard remaining around it? Use 3.14 for pi.

Hint →

Divide the perimeter by four to find the side length of the square first. For the circle, remember that radius is half the diameter.

Answer →

1600 square meters for the square, and 1521.5 square meters remaining.

Part 1: The square has a perimeter of 160, so each side is 40. Area is 40 squared, which is 1600. Part 2: The fountain's radius is 5. Area of a circle is pi times radius squared. 3.14 times 25 is 78.5. Subtracting 78.5 from 1600 leaves 1521.5.

From the thread

saevarom

I run the count at one of the constituencies. We have about 42.000 people on the electoral register and we'll probably see about 75-80% turnout, so we're going to be counting about 33.000 votes with about 30-35 people on staff, 24 of them doing the actual counting. For the interest of the HN crowd, here is the technical report: You'll find that the Icelandic voting system is very low tech. The electoral registry for each local precinct is printed out and each voter is marked as having voted by hand. Three copies are made and three independent staffers make their own marks in their copy. These registers then come with the ballots to the count site to possibly exclude absentee ballots. We use a paper ballot, marked using a pencil. Votes are counted by hand using a system of double counting and verification. Numbers are checked and cross checked throughout the process. The count is added and checked in an Excel sheet and then reported by telephone or sometimes live broadcast to the media. Official reporting is usually via custom reports in Excel to the national electoral commission.

seydorreply

Low tech enables multiple interested parties to physically oversee the electoral process. That is a good thing.

s_devreply

The reason low tech works for something as important as voting is because it's expensive and inefficient. That expense and inefficiency is what protects it from fraud. You would need enormous sums of money to overturn an election like this even for a place as small as Iceland and even then, such amount of money wouldn't guarantee the result of future referendums.

iamflimflam1reply

Not much different from the UK to be honest. The constituency and first past the post system means counting is a nicely decoupled distributed system. Each constituency declares the local winner and the only outside impact is the total count of MPs for each party. I would prefer something more complex that does give some of the smaller parties some MPs but no one seems particularly keen to change things.

inglor_czreply

We count votes in Czechia in very similar way and IMHO this beats any digital system by a mile. You don't have be an IT person to verify the count and trust the process, and the results for the entire country of 11 million are usually known in 3-4 hours.

ekidd

> The No campaign fears losing control over Iceland's prized fishing grounds under the EU's Common Fisheries Policy and has vowed never to share the country's waters with anyone. Brussels has indicated Iceland could earn some kind of exemption, but it is considered potentially the biggest obstacle to any agreement. As noted in the article, Iceland's fishing industry makes up something like 40% of their exports. Which is the only way a tiny island nation can afford essential imports. On top of that, Icelandic fishing is apparently carefully regulated to prevent overfishing, with the result that it is supposedly one of the few truly profitable fisheries in the world. So maintaining control over their own fisheries may be a survival-level issue for Iceland. On the other hand, they have a tiny population, no ability to defend themselves, and membership in the fraying NATO military alliance. So I can see why they're debating this issue. Aligning themselves more tightly with Europe doesn't fix their problems, but I can also see not wanting to go it alone.

glimshereply

They export good music too (Bjork and Sigur Ros)!

sscaryterryreply

The best part: https://en.wikipedia.org/wiki/Eurovision_Song_Contest:_The_S... (Edit: This is obviously a joke!)

AlotOfReadingreply

Why should it be a joke? Fire saga is beloved in Húsavík. My sister has this whole tour she does when we bring friends up.

Semaphorreply

Including a very healthy metal scene, Martröð (50 % icelandic) and Nexion were recent favorites of mine.

SyneRyderreply

Gus Gus too. And some of us still remember Greta & Jónsi, even out here in Australia.

lukanreply

Iceland doesn't even have a military at all.

boxedreply

And yet they defeated the UK navy several times :P

ben_wreply

IIRC this was done by threatening to leave NATO, which put pressure on the USA who really didn't want to lose a perfect base location to help monitor/contain the USSR, who in turn put pressure on the UK. But this is from memory; and given the sources I read may have been politicised (I'm British by birth), take with a pinch of salt.

Y-barreply

They can easily implement a large no-fly zone across Europe, as demonstrated with Eyjafjallajökull in 2010.

boxedreply

It's an existential threat not just to Icelandic economy but probably to the entire Atlantic ecosystem as a whole.

beloch

The EU needs to develop an associate option for close allies unwilling to become full members. Canada, Iceland, and others are very closely aligned with the EU in many respects and all could benefit from closer relationships. Ask a Canadian if they'd favour joining the EU and they'd probably be all for it, until you start talking about what that means. Fully joining the EU involves giving up a lot of sovereignty to a body that's too distant from places like Canada or Iceland. These countries are right to fear that their own concerns would be forgotten or ignored if they were to join the EU. Just as different structures are needed when a company grows past a certain size, the EU needs new structures if it's going to grow its influence beyond continental Europe. A middle ground needs to exist. It also needs to be very clear that associates, close allies, or whatnot are not inferior to full members, just less dependent on Brussels by choice because that's what makes sense for them.

anal_reactorreply

We already have associate options - not joining Schengen, not adopting Euro. Anything more than that results in a horrible mess of legal agreements. Imagine that suddenly each US state gets more sovereignty and they just... make different decisions. California adopts Spanish as second official language, New York adopts metric system, Alabama permits sister/brother marriages, Utah declares Mormonism as official religion, and Wyoming stops contributing to US army. How are you supposed to do business in these conditions.

unixfgreply

It's more of a patchwork here than you seem to think. Driving through DC with ammunition or mailing mifepristone to TX, as examples. CA would have to amend their constitution, but could pick a second language. NY already prefers metric AL could probably allow those marriages, they're not big on allowing marriages though. UT would be stopped by the establishment clause. https://law.justia.com/constitution/california/article-iii/s... https://www.nysenate.gov/legislation/laws/AGM/177

jankubicareply

Yeah, this already exists in various forms.

bradley13

There are two main problems with the EU, both political. First, the EU Commission has way too much power, and honestly wants to play dictator. Look at ChatControl: Voted down more than once, but the Commission just repackages and resubmits, with some dirty backdoor tricks, until they get it through. Parliament is not allowed to propose legislation, only the Commission. Second, regulatory hell. The continual avalanche of regulations out of Brussels cannot be overstated. People make fun of the bottle caps, but those are absolutely illustrative of how literally everything is over-regulated. This is one of the main reasons that European industry is struggling.

surgical_firereply

The EU commission is just a way for national governments to launder their authoritarian desires through the EU. The commissioner for each country is choser by national governments. It's them pushing that shit. The main problem with it, is that when you vote on national elections, typically you are concerned with local issues. Housing, healthcare, transportation, employment, immigration, etc. Nobody is really thinking "who will this party behave on the EU level?" Then they do shit like this and hide behind "we have to comply with the EU".

fidotronreply

> The EU commission is just a way for national governments to launder their authoritarian desires through the EU. The commissioner for each country is choser by national governments. It's them pushing that shit. You're both right. The problem is it's the commission that sets the agenda for the parliament. If you inverted it and had the elected parliament set the agenda for the national government appointed commission it would be far more likely to reflect the will of the people, which is precisely why it is the way it is. Europe still hasn't learned the lessons of the French revolution, and is destined to keep repeating it.

surgical_firereply

I am as pro-EU as it gets, and I really would prefer if it worked as a true parliament. Instead of commissioners chosen by national governments and a head commissioner, a premier minister chosen by the parliament, and a commission that would only have the power to approve of veto laws that come from the parliament. Hell, the parliament should even have the power to override vetoes. The way the EU is composed right now, it is designed to be kneecapped by the national governments every time.

Ylpertnodireply

> bottle caps. Oh, the humanity! And the lack of bottle tops in the streets. I'd move to the US, but your lack of regulations scares the piss out of me. And your cops, and your electoral system, overall crime stats.... Your first paragraph rings true.

From the thread

simonebrunozzi

Nice article by Stephen Diehl - I really liked his latest on crypto. I noticed that in certain sections there seems to be a heavy use of LLM-assisted writing, such as this one: > The platforms did not invent fear, greed, loneliness, or status anxiety. They industrialised their extraction. Their recommendation systems are vast reinforcement-learning loops that continuously experiment on human weakness. Each objective is a moving composite of high-dimensional signals for attention, retention, and conversion, dispersed across models, metrics, tests, and feedback systems. The subject cannot see the experiment. The operator cannot fully explain it. The regulator can barely comprehend it. The loop knows only that one stimulus keeps a person scrolling while another lets them leave. Calm accuracy loses. Threat, transgression, humiliation, and impossible promises win. The resulting social damage appears nowhere in the objective function. It arrives as an externality. Not sure what you think of it. In my view, his ideas are nice and interesting, and I would rather read less verbose, less "perfect" blog posts, and reduce or remove the use of LLMs.

maqpreply

Same. Every time I see LLM has contributed to the work I feel it greatly dilutes it. I'll push through if the topic is important and there's good points. But e.g. on YT, a sloppy script that barely stands on its own becomes a hard no on the very first "it's not X, it's Y".

ChiperSoftreply

That paragraph jumped out to me, as well. Thats also when the writing got very long winded and seemed to be repeating itself.

MattyRadreply

That does indeed seem like LLM text to me, which makes a depressing article even morning depressing. Inauthentic and hypocritical. I'd like to give the author the benefit of the doubt, but since we've set all ourselves on being cynical....

stillpointlab

One thing people go through as they age is an elevation of cynicism and general misanthropy. After decades of being disappointed when naive idealism and speculative optimism fail, one can just get down on it all. It is really hard not to succumb to this. It requires active effort and sometimes even a willful ignorance. "turn a blind eye" is the idiom here. Monty Python's "Always look on the bright side of life" kind of thing. Mr Rogers said it: look for the helpers. They may be few, but they are there. Find them and focus on them.

eventualcompreply

Thanks for commenting and acknowledging that it's challenging. My mates and myself have definitely fallen into this. Reminds me of that hackernews post about how doomerism is a siren's song for genZ, forget where it is.

noir_lordreply

I think that's why historical education is so important (or just liking history as in my case), looked through the lens of history all of the current shit as bad as it is, is really nothing all that new either, the only constant is change after all. As the aphorism goes: History may not repeat but it sure does rhyme.

watwutreply

The knowledge that 1932 was kinda similar is not comforting at all. Or whatever run up to historical horrible events we talk about. I genuinely dont like what history sais follows after times like this.

AlexB138reply

You're absolutely correct that history helps you see that none of this is new. However, is it better to be the person who sees the waterfall their boat is inevitably going to go over, or the person completely ignorant of it until it happens?

ChickeNESreply

> However, is it better to be the person who sees the waterfall their boat is inevitably going to go over, or the person completely ignorant of it until it happens? Neither, it's better to be the person who realizes we're in for a bright future no matter what doomers think.

mindwokreply

If we look at history, humanity has been through some pretty crazy shit and we haven't gone over the waterfall yet.

JumpCrisscrossreply

> see that none of this is new But this is. This sort of violent partisanship that self ensconces around identity versus issues is exactly what tears apart democracies and republics.

queenkjuulreply

I would kinda be with you, except, climate change. And like, we didn't have nukes when the last world war started. This time we will.

apireply

If you can’t help, it’s okay to look away. People sometimes get mad at this, like it’s callous or privileged, but what’s the difference between focusing on problems you can’t help and slowing down to stare at a car wreck? If you don’t slow down are you being insensitive? Obviously if you can help it’s different.

Glyptodon

So I think there is an oversight in this essay, and it's that society becoming predatory is much bigger than the Internet: it is the story of the decay of America, reflected and exacerbated by the internet, as it becomes a anti-democratic huckster society dominated by grifters, theocrats, and snake-oil sellers. Forget the internet, you can't get your home repaired, your car worked on, an appliance purchased, shop for food, or do anything else without wading through a corrupt morass of that mixes woo, subcultures, grift, and subscriptions. Maybe it's the hospital that wants to charge a $25 a day fee to make the hospital room TV grant access to basic cable channels instead of nothing. Maybe it's the elementary school district that mandates every physical interaction between students is called in to the police - like little Jenny shoved Johnny who skinned his knee because he teased her, call the cops, it's policy. Or the way that every customer service interaction is routed through endless AIs because it's cheaper than solving a problem. Or how it seems like nobody will do right by anyone else until lawyers get involved. Cut taxes year after year, while everything decays, we fail the younger generations, and everyone thinks you fake it till you make it because it's all zero sum... It's how America heads straight to banana republic, and it might be too late to turn.

From the thread

codethief

> Notably, Hy4 preview also contributed to its own development process, participating for the first time in the automated optimization of training methods, data strategies, evaluation frameworks, and low-level operators. The model proposed approaches, ran experiments, and iterated based on the results, with the resulting code, logs, and feedback feeding into subsequent rounds of exploration. This established an early-stage recursive self-improvement loop. This reminds me of one of the predictions from https://ai-2027.com/ . Only that there it's "OpenBrain" doing this, not the Chinese. And the authors of that paper were also slightly wrong about "Mid 2026: China Wakes Up": China woke up already a while ago. And: > But China is falling behind on AI algorithms due to their weaker models. The Chinese intelligence agencies—among the best in the world—double down on their plans to steal OpenBrain’s weights. No need to steal anything, they have already caught up. And then there's this prediction for February 2027: > Officials are most interested in its cyberwarfare capabilities: Agent-2 is “only” a little worse than the best human hackers I think we're past that point now, too…

bredrenreply

If the distillation "attacks" created useful inputs to open weight models, ai-2027 was directionally correct that the Chinese would find ways to extract IP from western firms. (Scaled account creation and grinding outputs etc is not a dramatic story element as spies, though!) Whether the distillation has constituted "attacks" or has or will meet the bar of "stealing" IP is not super interesting to me, though.

vlyanreply

the chutzpah of calling it an `attack` or `stealing` is super interesting tho.

haazreply

The idea that distillation is a significant contributor to the capabilities of the Chinese models is not true. Kimi K3 came out 2 weeks after Fable and uses a number of novel NN architecture innovations.

0xbadcafebeereply

The AI 2027 paper/website is exactly the same as random guesses from tech bros after a couple of beers telling you what they think the future will be. It has nothing to do with political theory, economic theory, game theory, or any other quasi-scientific or rigorous evaluation of real world events and predictable outcomes. It's just vibes. If they're wrong nobody will notice, if they're right people will call them geniuses.

try-workingreply

Just like how Windows 95 contributed to its own development process.

minimaxir

Hy4 apparently has ludicrous traction on OpenRouter already (https://openrouter.ai/tencent/hy4-preview), with trillions of tokens processed in a couple days: more than GLM 5.3 in a week. That said, it's relatively cheap with a 5% cache cost when everyone is still doing 10%/20% cache costs, so Hy4 may be more compelling.

cyanydeezreply

i'd be curious if openrouter is just being gamed by these publishers by paying for the exposure. wouldn't trust they dont do Capitalism like the rest of the AI field.

drob518reply

Of course they are. Of course they do. Nobody should be surprised by this.

tokaireply

>dont do Capitalism like the rest of the AI field Like lobbying the US president to harm their competitors?

realoreply

I would suggest "lobbying" is not the correct word to describe all the corruption going on in the current USA administration cesspool.

martinaldreply

I wrote about this a couple of weeks ago. It's actually often the biggest cost and it tends to be hidden away on most platforms! https://martinalderson.com/posts/watch-out-for-cache-read-co... Btw I still haven't came across any decent model that is <$0.01/MTok cache costs apart from deepseek thru their official API (even with the price increases). Seems like a bit of an opportunity for someone to take - drop cache read costs significantly.

dakollireply

That's because Deepseek invented the paradigm of prompt caching, they are the SOTA when it comes these techniques. Despite them open sourcing all their research, nobody beats them. edit: I do wish openrouter would let you sort providers by Cache Hit % and Cache cost. These are the only things that matter to me at this point when choosing a provider.

minimaxirreply

You can click the table headers to sort Ascending/Descending.

Bolwinreply

Cache hit % on openrouter is not a good metric, it's mainly driven by openrouter's own provider juggling than the providers themselves

orbital-decayreply

>Deepseek invented the paradigm of prompt caching Caching was always here, you don't need to do anything special to get it on a single user local backend running a base model or a chatbot in the first place. Among commercial providers, OpenAI adopted it in 4o first.

sievereply

For me, an average long session results in about 200-300M cached input, 4-800K input, 2-400K output. Mostly the lower bound. Output depends on how much the model thinks. There are two problems here: - cache hit pricing (both Muse Spark 1.2 Contributor and MiMo 2.5 are around the $0.002-3/M mark) - cache persistence time Muse Spark drops the cache in less than 5m. MiMo keeps it around for at least an hour based on my experience with whoever is serving it for OpenCode. This difference itself will inflate bills massively. A 500K token input repeatedly read by MS 1.2 for full input price 12 times an hour = $0.60. You would be expecting $0.012. So a 50x difference. Same thing on MiMo 2.5 is $0.018 because of longer cache times.

sourcecodeplzreply

even with the 5m cache, Muse Spark Contribs is still best bang for your buck for the intelligence you get. it is basically the old dsv4-flash prices, but even more smart.

Dinuxreply

Which explains why almost none of my request go though

fastball

I wish model providers would stop committing chart crimes in their releases. - if you're gonna order the rest of the bar chart by rank, order your model accordingly. - if you're gonna highlight a winner in a table of benchmarks, don't highlight your entire model row in the table. Etc etc

jimbob45reply

I wonder if this is being reinforced via LLM because they see every other modeler doing the same thing.

nullbioreply

I'd bet there's a correlation between benchmaxxing and chart crimes. Companies who try to deceive perceptions via the charts are more likely to cheat at the benchmarks too, I'm sure. That's assuming ill intent, of course - which is often the case for charts related to model releases, but not necessarily always the case.

petcat

> Tencent has released and open-sourced Tencent Hy4 preview, a next-generation large language model with 770B total parameters and 49B active parameters, and a context window exceeding 1M tokens. There are no open source models, at least not useful ones (yet) [0]. Open weight is not the same as open source. The current "open weight" models are just opaque binary blobs you can run on your own computer instead of through a web API. [0] https://allenai.org/ Imagine thinking that running a Photoshop binary on your own computer instead of through a SaaS web app means that it's "open source". Of course you think that's ridiculous.

mirekrusinreply

You can open source dataset without all the details how it was assembled. Models are lossy compressed datasets you can pick up and amend (fine tune / continue training / alter) according to license they were released under. Hy4 is released under OSI approved Apache License 2.0.

kennywinkerreply

Parent poster is technically right - open “source” implies the source used to make something is open. The model source is training data and code, not just weights. But the reality is, the weights are a useful artifact that you can use to create derivative works. So, dismissing it as a photoshop binary is as technically wrong as calling it open source.

Alpha3031reply

IIRC Nvidia claims to release enough data that it should be possible to fully reproduce Nemotron, so even if it's not as good as the current best models, GPT 5.1 or Opus 4.1.was still useful right? I guess it depends on what you wanted to do with them.

LtWorfreply

So windows is open source because the binaries are a lossy compression of the original source?

NitpickLawyerreply

Weights are not binary. A model is created at init time, with random values. After that, it is being modified using data. The key point is that the labs modify the models "as weights". That means that weights are the intended / preferred way of modifying a model. Which, coincidentally, matches the definition of source in Apache 2.0. There is no "higher level" place where editing takes place. It all happens in weight space. Through the license you get the same rights as the lab that created it: view, inspect, run, modify, re-release. That's it. That's the only thing a license can grant you. The rest is semantics, misunderstandings, and FUD. A model released under an open source license is open source. Training data is lab knowhow / IP. Which, historically, has never been required for any open source release.

petureply

Before we worry about source code, Microsoft doesn't grant me rights to modify/redistribute/sell copy of Windows I have.

villishreply

Countries that aren’t competitive need access to training datasets so that they may train their own similarly capable models and be sure of the inputs. Governments cannot blindly trust open weight models from China and the US.

From the thread

ihsw

Did it reduce catalytic converter theft?

baldeaglereply

It’s hard to say. As the economic got better, trade policies stabilized and the price of commodities and therefore the metals in a converter dropped in value. As the price of the metals went down, the thefts dropped nationally. Texas thefts seemed to have dropped inline with the national trend. Since the supply side issues are picking up again, the thefts are going up likewise.

p_j_wreply

If what you say is true it doesn’t sound hard to say at all, it seems clear that the cameras didn’t reduce the thefts at all.

floralhangnailreply

I doubt it helped much directly. The nationwide ring that was dismantled recently was busted partially due to surveillance from a pole mounted camera though. The article title initially baited my rage, but I can plausibly see how LPRC's could be used to track catalytic converter thieves. https://en.wikipedia.org/wiki/2020%E2%80%932022_catalytic_co...

CoffeeOnWritereply

(a pole mounted camera for which a warrant had been obtained...)

s1artibartfastreply

Jeez, in not normally cynical but the Japanese buyer faced no charges and this has bad optics: > Janice Ayala joined Dowa Metals & Mining America as their Chief Compliance Officer on March 3, 2023. During Ayala's 27-year tenure at the Department of Homeland Security, her most recent role before joining Dowa was as Director of Joint Task Force – Investigations. This agency oversaw the investigation into both DG Auto Parts and Dowa.

cosmicgadgetreply

Doesn't seem especially likely, does it? People report the thefts in unmonitored neighborhoods. The police... check the nearby intersection cameras for any vehicles passing between 11pm and 4am?

Gigachadreply

As much as I think the tech needs to be regulated to prevent abuse, if I was a driver in San Francisco I’d try to only ever park my car directly in view of a flock camera.

mannanj

So to fight abuse and overreach by your leaders, you can: - give up your car/vehicle propery - let an LLC take ownership - reside in the privacy protections of LLC and corporation I don't think any successful case has happened yet where flock et al have pierced the corporate veil to get address information, nor dug into insurance records to find who the LLC-license plates belong to.

carlobreply

The fact that corporations have more privacy rights than people is dystopian.

toofyreply

i don’t think that will help anything. surely flock and similar companies use facial recognition features? particularly at every stores parking lot that has these in it? and i’d be more shocked if the same stores weren’t already sharing their own inside cameras facial recognition directly with companies like flock already. and from what we’ve seen, these companies cameras are linked up with names from frequent shopper cards, cc data, what people buy, when they buy it, etc…

SwellJoe

Given how unpopular these cameras are, and how sloppy their security has been, I have come to assume there's boatloads of bribery happening. I can't come up with any other explanation for how willing police chiefs and city councilors and mayors are to get yelled at at city council meetings and how wantonly they're giving away the privacy of their constituents. I'm guessing that as scrutiny increases, we'll begin to hear about questionable, at best, sales tactics, if not outright bribery. Trips, sponsored speaking/training engagements that puts money in the pockets of cops and policymakers, etc. The whole thing stinks of it.

bigbinaryreply

It’s been proven (at least in the Minnesota flock conference) that flock pays politicians and police as private contractors and commissions them on any potential contracts they bring to flock

From the thread

livinglist

I scuba dive in Monterey, CA every other weekend, and just came back from a liveaboard in Galapagos in May. The water is much warmer this year, and it is very noticeable because usually in Galapagos, divers are expected to wear 8mm wetsuit or dry suit to be comfortable in the water, but this year even a 5mm made me feel too warm. In Monterey, water temp used to be around 11-13°C but this year it’s easily above 16°C, divers also reported more sightings of spiny lobsters which were rare in Monterey because water was too cold for them.

AndrewKemendo

You can’t get a more direct example of failure of collective action. I’m not sure the human species has the capacity to actually coordinate behavior at the scale needed, irrespective of any future warning etc…, to be able to combat this

esikichreply

People can't even think big enough to zipper merge properly. We are absolutely cooked.

BearOsoreply

We all reduced CFC use and stopped ozone depletion. The problem with climate change is that there is a huge incentive to not fix it, and the culprit is not just rich oil corporations. Many countries subsist solely on oil profits and have no other natural resources to sustain themselves, i.e. Saudi Arabia. Those countries have influence over others like the US. There are alternative options already available to replace fossil fuels for most things. In first-world countries, the only thing stopping a transition is the loss of corporate profits.

Gigachadreply

Ironically the war in Iran was probably one of the most effective ways to cut oil consumption.

iLemming

We're getting cooked. I'm already paying 10-times more for water usage than I spend on gas, and 5-times more on electricity bills - AC's running non-stop. That's crazy, because I live in the area where you can't even walk to buy a piece of bread - we drive all the time. I fully expect for businesses very soon to operate at night. Costco, Walmart, groceries, even libraries will be working all night in places like Arizona, Texas and some parts of California.

sarchertechreply

> I'm already paying 10-times more for water usage than I spend on gas Are you irrigating an enormous yard?

edoceoreply

Maybe they mean consumption is the same but price is up.

sarchertechreply

The only way I can read what they wrote to mean that is if they are saying water costs $30-40 a gallon.

iLemmingreply

No, but I have a pool and it's been drying up below the skimmers, I need to keep the water level or the filtration dies.

vachinareply

lol. What a first world problem to have. Sounds like a very solvable problem too (remove the pool then)

pvaldesreply

Maybe a sheet of plastic floating in the surface while you don't use it could help to reduce the evaporation.

folkravreply

Maybe this country focusing so much on keeping the gas cheap might have something to do with it.

slashdevreply

US emissions have declined 18% since 2005. Look at Asia.

NegativeLatencyreply

You mean the place where we buy all our crap from?

nyc640reply

Yes, let's look at "Asia" (as if that's some meaningful grouping in this context when the countries are so different culturally, geographically, and economically). Since 2005, a huge amount of US manufacturing and resource processing (both large emitters) have been outsourced (largely to Asia, and more specifically China). Our economy doesn't function without these imported emissions, yet they don't count as "US emissions". Even excluding these emissions, the US makes up 4% of the world population and emits 13% of world emissions. Asia makes up 60% of the world's population and emits 52% (more than 3x less per-capita than the US). If you actually consider that the countries in Asia are vastly diverse, you'll find the large majority of these emissions are from China (from said manufacturing and resource processing), and India for example emits far less than its share of the population. This also completely ignores that the effects of warming are from cumulative emissions (not just point-in-time emissions) because CO2 doesn't just disappear every year. As a share of historical emissions, Europe comes off far worse than Asia per capita, and the US even worse than Europe. Even despite this, at least China (and the rest of Asia to a lesser extent) is investing in clean energy. China produces > 80% of the world's solar panels and > 70% of the world's lithium batteries. Meanwhile, in the US the administration is doubling down on subsidies for fossil fuels and issuing stop-work orders on renewable energy projects already under construction. And because I know I have to give this disclaimer these days: I am no apologist for everything else wrong with China and their authoritarian government. But I can't bear this attitude of absolving ourselves of blame for climate change because some other continent has a bigger number than us.

Hikikomorireply

China built more solar in 2025 than the US has in total. Meanwhile trump ripping up wind and solar builds.

iLemmingreply

American gas is not cheap. In absolute-numbers-sense, it's comparable to the current prices in Russia. And by all means and for all the different reasons, one would expect gasoline to be enormously expensive there right now. Far more expensive than in the States. It doesn't seem to be the case.

Aurornisreply

> American gas is not cheap. Yes it is. I don’t know why you picked Russia to compare to. Russia’s government heavily intervenes in markets to keep prices low, including subsidizing refineries and banning exports at times. The fact that our prices are as low as theirs means our gas is cheap.

0n0n0m0uzreply

Libraries in SW Colorado were packed! Alot of people just using the AC since most houses dont even have it

iLemmingreply

A few years ago (before wildfires forced us to move), we were in the East Bay of San Francisco. East Bay gets unusually hot - the mountain range prevents cool winds from the ocean passing over. Summer was extreme, the AC couldn't keep up, we had to hose the outer walls every evening, just to cool it a bit. One day we woke up, had breakfast and the power line had failed. I quickly imagined what happens if it lasts more than a few hours, checked the heatmap, found an area that wasn't dark orange, and told everyone to pack - we're leaving. We headed towards Half Moon Bay. It's a little over an hour drive, but it took us another forty minutes to find a beach with a room for parking - entire coast was packed with the heat escapees like us. The beach was full of stinking kelp, there were tons of flies - they've tormented our stay, but we had little choice. We thought we were lucky to find a spot with not too many people, only to realize there was a carcass of a dead sea lion and it was dead-cat-stenchy. At least, being next to the ocean offered some relief from the heat. Later that afternoon, our neighbors called us to tell us that they were sitting in their minivan, at an empty parking lot. The power was still out, and to have a dignified dinner, they've grabbed some food and convened in the car, which had a working AC. But they didn't want it to look weird, so they drove away from their street, away from curious neighbors and friends. Next summer wildfires came back. This time bigger than ever. Dealing with both smoke and COVID restrictions together felt like too much, and we decided to move. We headed East, our first stop was Barstow, CA - we decided to rest a night. It's pretty far from where the actual fires were raging, yet when we woke up, we were greeted with what looked like snow at first. It was ash. Our car was covered with it. We kept driving. The sky was black from the smoke. It only started clearing out when we crossed into New Mexico. We were all terrified to realize the scale of that disastrous event. Which has now become "a seasonal climate occurrence" - it's almost guaranteed that each year something will massively burn somewhere in the West Coast.

v2026.08.30

Weather

World Clocks

Markets

v2026.08.30

Weather

World Clocks

Markets

The Debate

Did Neandertals independently develop symbolic material culture before the arrival of Homo sapiens in Europe?

Does the predictive success of idealized scientific models justify believing that their unobservable entities actually exist?

Can transgenerational epigenetic inheritance drive permanent adaptive evolution in complex animals without changing DNA sequences?

From the Labs

Sony Music and Warner Chappell Sue Anthropic Over Copyrighted Works

Hugging Face Adds Hindi and Indian English Sets to Open ASR Leaderboard

Google Research introduces planetary prediction engine for global models

Hollywood celebrities enter microdrama market through apps and investments

Google DeepMind releases Gemini Omni 1.1 Flash with enhanced video controls

Therabody launches Theragun Sense massage gun with guided LCD routines

Google DeepMind pilots double-blind AI evaluations using hardware enclaves

XREAL a01 smart glasses deliver vivid displays alongside thermal and utility trade-offs

Google Research introduces GlucoFM for continuous glucose monitoring

Lambda secures $1B in debt to purchase Nvidia GPUs for Microsoft

European Tech Leaders Debate AI Sovereignty at Copenhagen TechBBQ

Vijay Pande launches VZVC to make concentrated bets in AI biotech

Nvidia shifts focus to data orchestration and whole-system infrastructure

Chinese automakers enter humanoid robotics with massive private funding rounds

Anthropic Details Automated AI Researchers for Alignment Post-Training

Brave Browser Adds Native Email Aliases for Privacy Protection

Google DeepMind launches Gemini 3.5 Transcribe speech-to-text model

Hacker News

Debian votes to allow "responsible use of generative AI"

Iceland votes on whether to restart talks on joining EU

The internet is kind of a predatory cesspit now

Hy4 preview

Lawmakers added $1 to car insurance policies. That money paid for Flock cameras

Highest-ever ocean temperature measured as powerful El Niño forms

The Rabbit Hole

Deaths in 2026

Research

Sophistication in GenAI Use: Field Evidence from a Large Firm

Social Network Structure, Wealth, and Wealth Inequality Across Cultures

AI worsens climate change, integrated assessment shows

From the Bookshelf

The Death and Life of Great American Cities

Eyes on the Street

Contrarian Corner

Polyamorous relationships produce net negative outcomes for most individuals who attempt them.

The Supreme Court has largely erased the press from its First Amendment rulings over recent decades.

Fewer than half of Republicans hold a favorable private opinion of capitalism.

Photograph

On This Day

Games

Puzzles

The Three Storage Units

The Salary Hike

The Architect's Walk

The Track

Cariño

Quote

Word

Komorebi

Pareidolia