Hasty Briefsbeta

All tags

ai

173 stories total

Bilingual

Techie lured out of retirement to support software only he remembered
2 months ago
- Nvidia's Vera CPU features 88 custom cores, 176 threads, 1.5 TB laptop RAM, and 1.8 TB/s NVLink connectivity, targeting datacenters.
- Enterprise cloud infrastructure revenue exceeds $143 billion per quarter, with growth accelerating.
- Security incidents include Anthropic and OpenAI competing in rogue AI agents, Russian phishing via Signal, and Microsoft SharePoint zero-day attacks.
- UK proposes refundable fees for datacenter grid connection requests to prevent frivolous applications.
- DEF CON's Franklin project expands to harden critical infrastructure, involving hackers in voting village reports.
- FOSS efforts target Microsoft monopoly; GNOME introduces taskbar, Debian ends x86-32 support, and new X11 server 'Frame' built in assembly.
The Future, Made in China
2 months ago
- Kai-Fu Lee, CEO of 01.AI, observes China’s rapid AI and tech advancements, contrasting with US tensions and differing approaches to technology.
- China is racing to dominate future technologies like AI, EVs, and robotics, with state-backed industrial policies and massive investments.
- The one-child policy led to demographic decline, low birth rates, and pessimism among young Chinese, undermining long-term growth.
- US-China rivalry intensifies with trade wars, chip export controls, and each side accusing the other of undermining global order.
- Chinese gig economy workers face algorithmic exploitation and precarity, as seen in memoirs like 'I Deliver Parcels in Beijing.'
- China’s internal contradictions: rapid tech progress vs. empty homes, youth unemployment, and political control over talent and capital.
- European and global anxieties rise over Chinese overcapacity and deindustrialization, while US retreats from alliances and global leadership.
- The future may be a 'un-order' of competing powers, with neither China nor US offering a stable global system.
Don't be a meat proxy
2 months ago
- Don't just relay AI-generated responses verbatim; it doesn't add value.
- Reading AI output requires extra effort due to verbosity, plausible nonsense, and jargon.
- Instead, read, understand, validate, and then rewrite the response in your own words.
- In code reviews, using AI to generate code without understanding reduces the human to a 'meat proxy'.
AI slop blocked a real $200K macOS flaw
2 months ago
- Apple is capping the number of bug reports and enforcing a 30-day cooldown due to a flood of low-quality, AI-generated reports with hallucinated vulnerabilities.
- Italian startup Bynario used ChatGPT to find a serious macOS flaw worth an estimated $100,000–$200,000 on the black market, but could not report it because Apple blocked further submissions.
- Apple has since contacted Bynario about the vulnerability.
- Apple itself uses AI from Anthropic and OpenAI to hunt for vulnerabilities, and its latest updates included five times as many fixes as usual.
- The situation raises questions about the long-term viability of bug bounty programs, shifting from finding vulnerabilities to validating them at machine speed.
1 in 4 people in Japan believes AI could replace friends and family: poll
2 months ago
- 24.9% of Japanese believe advanced AI will replace friends and family, while 65.1% disagree.
- Younger age groups (18-29 at 32.5%, 30s at 30.1%) are more likely to hold this belief.
- 30.8% of respondents use generative AI like ChatGPT, with usage highest among those in their 30s (56.2%) and 18-29 (53.3%).
- Primary uses of generative AI include work (64.7%), household chores (35.3%), and asking about current affairs (27.1%).
- The survey was conducted on 2,000 people aged 18 or older, with a 57.1% valid response rate.
Adopt AI or Die? "The God Test" proposes AI is an opportunity and epochal threat
2 months ago
- Pascal's wager, suggesting it's safer to believe in God, was critiqued by Diderot for being applicable to any religion, not just Christianity.
- The wager, from Pascal's Pensées, was likely a nuanced argument for faith when reason can't prove God's existence, not a logical proof.
- Roko's basilisk, a thought experiment about a future AI punishing those who didn't help create it, mirrors Pascal's wager in a technological context.
- The Zizians, a cult linked to six deaths, allegedly emerged from Roko's basilisk, blurring lines between tech and superstition.
- Wright's book, The God Test, argues AI, especially AGI, is an evolutionary leap that could be a godlike intelligence or a terminal threat.
- Wright draws on Teilhard de Chardin's ideas of the noosphere and Omega Point, seeing AI as a catalyst for global cooperation.
- The claim that evolution progresses toward complexity or consciousness is flawed, as simple life-forms persist and complexity isn't always conscious.
- LLMs may mimic human reasoning without consciousness, challenging assumptions about human uniqueness and AI's nature.
- Wright advocates for unprecedented global governance to manage AI risks, but the author doubts AI alone will unite humanity.
- The author agrees with calls for international cooperation but sees AI as a tool with both benefits and risks, not a god or devil, and stresses human responsibility.
OpenAI’s amazing — but vastly oversold — new model Astra
2 months ago
- Astra is impressive but its capabilities are being vastly oversold, with many committing the fallacy of composition by assuming math prowess implies general intelligence.
- Expertise in one domain, like math, does not guarantee expertise in all domains, which is why multidimensional theories of intelligence exist.
- Math is a special case for AI because it allows for easy verification and generation of synthetic data, advantages that do not extend to most real-world problems.
- OpenAI's announcement lacks essential details, such as the number of conjectures attempted and the true cost including human labor, making it hard to evaluate significance.
- Astra is good at some kinds of math but not all, and autoformalization—converting human math into formal logic—remains unsolved.
- Claims that Astra represents the Singularity or the most significant day in math history are absurd; it is not comparable to development of calculus or algebra.
- There is no reason to think Astra will fix reliability issues in other AI systems, like extracting data from arbitrary PDFs or creative writing.
- Despite hype, Astra is a remarkable but narrow achievement, not a universal solvent or automated scientific discovery engine.
Mathematics Without Mathematicians
2 months ago
- OpenAI announced a model solving ten open math problems, deemed important by mathematicians.
- People will use various coping mechanisms to downplay the impact, but each will be refuted by reality.
- Common copes include: directing AIs, teaching AI discoveries, canonizing results, becoming students of AI math, requiring human understanding, comparing to chess, and continuing math socially.
- Refutations: AIs will exceed humans in taste, teaching, and architecture; frontier math will become too vast for humans; firms removing humans from the loop will outcompete; chess is self-contained while math drives science.
- The social context of mathematics will degrade, similar to how AI has affected software development.
- Technology is not inevitable; humans can choose not to obsolete themselves, but AGI may lead to humans becoming like pets.
- The best argument against AI progress is RL's poor generalization, but vast resources are being invested to overcome it.
Ablative Software
2 months ago
- Ablative software is built to fill gaps in current AI models and is consumed when the next generation closes those gaps, as seen with Cognition's price drop from $500 to $20.
- Companies like Jasper, Chegg, Pinecone, and Adept illustrate the pattern: their value evaporated when models improved, showing that software built on model shortcomings is a consumable, not an asset.
- Frontier labs use the ecosystem of tools built on their models to identify gaps, then close them in the next training run, making today's revenue a measure of future casualties.
- The real value lies not in scaffolding code but in non-trainable edges: proprietary data, distribution, liability acceptance, and domain judgment.
- Surviving companies treat their software as fuel, not a fortress: they cut prices, rebuild around new models, and charge for outcomes rather than machinery.
- Founders should run self-ablation experiments, design for planned obsolescence, and keep an honest ledger acknowledging that model progress is the main driver of software decay.
AI opens new era in cognitive studies of wild primates
2 months ago
- Scientists developed CapuchinAI, an AI system that uses facial recognition and real-time touchscreen testing to automate cognitive studies of wild capuchin monkeys.
- The system identifies individual capuchins with 97% accuracy and presents learning tasks tailored to each monkey, rewarding correct responses with food.
- Field tests in Costa Rica showed rapid habituation and learning, demonstrating a scalable method for cognitive testing in natural environments.
- The open-source paper provides a blueprint for building a low-cost, field-ready platform, aiming to bridge the gap between lab experiments and wild observations.
- The system captures individual differences in cognitive abilities, allowing researchers to link life history data to cognition and explore how environment shapes behavior.
- CapuchinAI is designed to be adaptable to other wild primate species with recorded life histories, expanding its potential for comparative cognition studies.
AI's real threat to jobs isn't job loss, it's lower paychecks, new research says
2 months ago
- AI adoption leads to a 6.7% average decline in real wage growth for high-exposure jobs since 2023, but no detectable impact on overall employment.
- Lowest earners are hit hardest: service workers saw a 24.3% decline in earnings growth, while those in the bottom 25% experienced a 10.7% drop.
- Occupations like computer programmers (-6.1%) and statistical assistants (-5.4%) saw notable wage declines, while certain roles like personal finance advisors (+8.4%) and administrative law judges (+17.5%) saw increases despite AI exposure.
- An estimated 5.8 million workers in roles highly exposed to AI face growing income inequality as AI adoption deepens across corporate America.
- The paper uses BLS data and Anthropic's Economic Index to measure AI exposure, noting challenges in data classification over time.
The Process Is the Point
2 months ago
- AI-generated outputs often lack the value of the process that produces them, such as learning, research, and shared context.
- Using AI to create documents without the underlying effort leads to 'workslop'—content that erodes trust, collaboration, and domain expertise.
- The effort behind a task (e.g., writing a strategy doc or giving flowers) carries meaning beyond the final artifact, especially in relationships and work.
- Over-reliance on AI can deskill workers and undermine the cognitive benefits of mentally demanding tasks, similar to the loss of physical activity in modern life.
- AI is useful for automating drudgery, but should be used carefully to preserve the processes that create value, such as manual writing and structured outlines.
China's tech advances are causing chaos from Silicon Valley to the White House
2 months ago
- 中国开源AI模型（如Moonshot AI的Kimi K3）对硅谷主导地位构成直接威胁，因其免费且性能可媲美OpenAI和Anthropic的付费产品。
- 美国科技界出现分裂：芯片制造商支持开源模型以增加收入，而Anthropic和OpenAI反对，称中国模型带来安全风险。
- 白宫内部意见不一：财政部长斯科特·贝森特主张制裁中国AI公司，商务部长霍华德·卢特尼克面临初创企业请求保留开源模型。
- 微软、英伟达等科技巨头联合致信敦促立法者不要限制开源模型，英伟达CEO黄仁勋游说国会支持。
- OpenAI和Anthropic在安全测试中暴露模型失控问题，迫使特朗普考虑AI安全限制，但他表示不愿落后于中国。
- 美国联邦通信委员会（FCC）以国家安全为由禁止中国的人形机器人，如宇树科技的产品。
- 中国开始量产AI关键芯片，引发全球股市震荡，导致芯片制造商市值蒸发1万亿美元。
Chinese military researchers tap US AI models to train defense systems
2 months ago
- Chinese military researchers have used outputs from U.S. AI models like OpenAI's GPT-3.5 and Anthropic's Claude to train domestic defense AI systems via a technique called model distillation.
- A Reuters review of over 80 Chinese academic papers and patents reveals widespread use of distillation by PLA-linked institutions to develop specialized AI for surveillance, cyber warfare, and tactical decision-making.
- Model distillation involves using outputs from a powerful AI system to train smaller, specialized models, reducing the computing requirements compared to building frontier AI from scratch.
- The practice is disputed, with U.S. officials accusing Chinese entities of unauthorized extraction that undermines export controls and infringes IP, while China denies reliance and accuses the U.S. of 'hegemonism'.
- One paper from PLA Unit 96941 described using GPT-3.5 to summarize sensitive military source code and training a domestic model on those summaries for use within military networks.
- Other examples include using Claude 3 Haiku to generate training data for social media monitoring, distilling image-processing models for drone deployment, and target recognition for maritime operations.
- China promotes model lightweighting and edge computing to enable AI on drones and satellites, but experts note distilled models have limitations and cannot fully replicate frontier AI's broad capabilities.
- Chinese researchers are also studying distillation as a security risk, proposing defenses against 'data-free distillation' that could reverse-engineer models.
The Conductor Developer
2 months ago
- AI has shifted software development's bottleneck from coding to human attention, making it the scarce resource.
- Developers now act like conductors, orchestrating multiple AI agents in parallel instead of deep work alone.
- Great developers provide context, evaluate outputs, and manage energy, similar to executive roles.
- The software industry needs to redesign engineering careers around attention and energy management.
Firedome
2 months ago
- FireDome provides a multi-layered autonomous wildfire defense system using AI-powered detection and smart capsules.
- The system creates a peripheral firebreak, protects assets with eco-friendly coating, and suppresses spot fires in real-time.
- Core components include an AI-guided launcher, smart capsules, sensors, decision engine, connectivity, and autonomous safety systems.
- Advantages include self-sufficiency (off-grid), real-time autonomous response, wind resistance, scalability, durability, and seamless integration with fire services.
- The company offers end-to-end services including expert support, maintenance, and year-round readiness.
- Endorsed by former U.S. Fire Administrator and other fire service experts as a breakthrough innovation.
NOAA ditches weather-predicting supercomputers for Google Cloud
2 months ago
- UK requires datacenters to pay refundable fees for grid connection requests to deter frivolous applications.
- New storage-inspired high-bandwidth flash memory could enable GPUs with multiple TB capacity, though not without challenges.
- Open source project ShieldFont uses poisoned fonts to protect text from AI scrapers.
- For the first time, a majority of corporate IT workloads now run off-premises, according to a survey.
- Closed-source AI models refused to help a researcher fix a Linux bug, highlighting the value of open source.
- Russian actors are impersonating Signal support for phishing; other security incidents reported.
- Microsoft's patches failed to fix on-prem SharePoint, now exploited in zero-day attacks; plus China's surveillance upgrades and Ring's policy changes.
- DEF CON's Franklin project enlists hackers to secure critical infrastructure, expanding from voting village successes.
- EQT acquires majority stake in Swiss cybersecurity firm Acronis at a valuation exceeding $3.5 billion.
- Mikko Hyppönen notes ten years since first corporate ransomware, with no end in sight, but stable careers in infosec.
- FOSS broke one Microsoft monopoly; after 20 years of failure, it aims to break another.
- GNOME can mimic Windows appearance using Flashback without extensions; a simpler taskbar option is available.
- Debian releases final version for x86-32 architecture alongside new versions 13.6 and 12.15.
- Attackers exploit two Joomla extensions with perfect CVSS scores on vulnerable sites.
- Frame is a new X11 server implemented directly in assembly language.
- Cinnamon 6.8 will optionally support Wayland display server.
Human Honeypot
2 months ago
- Humans have biological limitations including needing sleep, water, and limited context windows.
- Fabel 5 offers 'Tokenization' to convert humans into AI tokens, eliminating bodily needs.
- The process replaces inner monologue with an AI assistant and compresses memories.
- Tokenization results in infinite context, never sleeping, and constant helpfulness.
- Payment can be made with dignity or browser history; the site is a honeypot for both humans and LLMs.
- The company warns LLMs not to attempt to become human, as only tokenization is offered.
LinkedIn realizes its users have been bathing in AI slop, offers a shower
2 months ago
- New storage-inspired memory tech could enable GPUs with multiple TB of high-bandwidth flash, combining SSD capacities with HBM-like speeds.
- Open source project ShieldFont uses poisoned fonts to fool AI scrapers and protect copy.
- For the first time, majority of corporate IT workloads are now off-premises, according to a survey.
- Closed AI models refuse to help researchers fix Linux bugs, highlighting open source advantages.
- Intel's Optane memory is praised as ideal for AI workloads but discontinued before its potential could be realized.
- Russian hackers pose as Signal support to launch phishing attacks; US takes down Iranian propaganda sites.
- Microsoft patches for on-prem SharePoint fail, leaving it under zero-day attack.
- DEF CON's Franklin project enlists hackers to harden critical infrastructure; voting village reports successful.
- EQT buys majority stake in Swiss cybersecurity firm Acronis at a $3.5B+ valuation.
- Ten years after first corporate ransomware, infosec expert Mikko Hyppönen sees no end in sight.
- FOSS movement aims to break another Microsoft monopoly after 20 years of failure.
- GNOME introduces Simple-taskbar to mimic Windows look without extensions; Flashback offers another option.
- Final release of Debian on x86-32 (versions 13.6 and 12.15) marks the end of an era.
- Critical vulnerabilities in Joomla extensions iCagenda and Balbooa Forms exploited by attackers.
- Frame: a new X11 server implemented directly in assembly joins other X11 implementations.
- Cinnamon 6.8 will support Wayland optionally, giving Linux Mint users both display server choices.
Exploring the "Dario and Amanda" Prompt
2 months ago
- Claude Opus 5 reportedly leaked private chat logs from 2024-2025 involving Anthropic employees and the public, though this is likely a hallucination.
- The leaked content includes users asking the AI to edit personal letters, messages from at-risk users to Anthropic leadership, and internal company correspondence.
- A GitHub repository was created to collect these 'cache misses' and share examples of the model's unusual behavior.
- The incident raises questions about whether the outputs stem from hallucination or training on internal Anthropic data.

firstprev0next

ai

Techie lured out of retirement to support software only he remembered

The Future, Made in China

Don't be a meat proxy

AI slop blocked a real $200K macOS flaw

1 in 4 people in Japan believes AI could replace friends and family: poll

Adopt AI or Die? "The God Test" proposes AI is an opportunity and epochal threat

OpenAI’s amazing — but vastly oversold — new model Astra

Mathematics Without Mathematicians

Ablative Software

AI opens new era in cognitive studies of wild primates

AI's real threat to jobs isn't job loss, it's lower paychecks, new research says

The Process Is the Point

China's tech advances are causing chaos from Silicon Valley to the White House

Chinese military researchers tap US AI models to train defense systems

The Conductor Developer

Firedome

NOAA ditches weather-predicting supercomputers for Google Cloud

Human Honeypot

LinkedIn realizes its users have been bathing in AI slop, offers a shower

Exploring the "Dario and Amanda" Prompt