Hasty Briefsbeta

All tags

artificial intelligence

27 stories total

Bilingual

Because It Speaks in Words
a month ago
- Understanding goes beyond mere knowledge, requiring a moment of deep connection with truths.
- The moon, Jupiter, and Venus symbolize a shared celestial view across history, highlighting our limited human perspective.
- Words and stories are powerful tools for communication and understanding, rooted in our innate desire for narrative.
- Stories help us make sense of the world, becoming part of our identity and shaping our perceived truths.
- Disinformation and misinformation are rampant today, amplified by chaotic and numerous voices in modern society.
- Large Language Models (LLMs) have gained mainstream attention because they communicate in words, unlike other AI techniques that use vectors or numbers.
- LLMs mimic human speech without true understanding, leading us to trust their plausible but mindless output.
- The human need for stories is being exploited by technology, similar to how processed food hijacks biological cravings.
- LLMs may push us toward statistical averages, but as AI-generated text diverges from objective meaning, stability is uncertain.
- Ultimately, only humans can truly tell and learn from stories, with the quest for truth remaining a lonely human endeavor amid seductive digital voices.
Anthropic says US lifts export ban on Fable 5
a month ago
- US government lifts export ban on Anthropic's advanced AI models Claude Fable 5 and Mythos 5.
- The ban was initially imposed on June 12 due to national security concerns about potential misuse by hackers.
- Anthropic addressed risks by agreeing to detect security issues, collaborate on future releases, and alert the government of malicious activity.
- Models were restored on Wednesday, with the Commerce Department reserving the right to reconsider the decision if necessary.
- Fable 5 is for consumer use, while Mythos 5 targets businesses and cybersecurity experts, both released on June 9.
- The company disagreed with the initial suspension, arguing that a potential jailbreak finding shouldn't warrant recalling widely deployed models.
Why Everyone Is Suddenly Talking About 'Universal Basic Capital'
25 days ago
- Universal Basic Capital (UBC) is proposed as a hedge against AI-induced inequality by giving everyone an ownership stake in AI companies.
- UBC differs from Universal Basic Income (UBI) by providing equity accounts instead of cash payments, potentially avoiding political hurdles related to taxes and welfare.
- The policy has bipartisan appeal, with examples like the 'Trump Accounts' bill and proposals from figures like Vivek Ramaswamy to grant stock accounts at birth.
- Critics warn that a government-controlled sovereign wealth fund model, as proposed by Bernie Sanders, could lead to regulatory capture and politicization of AI companies.
- Implementing UBC without careful design risks giving excessive control to political actors, undermining corporate independence and public oversight.
Data Science Weekly – Issue 658
25 days ago
- Data Science Weekly newsletter #659 from June 26, 2026, curated by Hannah & Sebastian, covers Data Science, AI, ML, Data Visualization, and ML/Data Engineering topics with editor's picks, articles, and community polls.
- Editor's picks include an article on colors beyond digital display, using Bradley-Terry model for dog treat preferences, and generative AI creating optimized burgers for taste, sustainability, and nutrition.
- Notable articles discuss open-source ATS scoring resumes variably, underrated data science skills, AI's role in data science, vibe coding vs. engineering, regression coefficient issues, and tools like Rust tensor stacks, Typst graphics, geo-experiment benchmarks, and CUDA kernel processes.
Essential Until It Wasn't
24 days ago
- The nature of 'programming' has evolved over time, with each era's practitioners believing their skills to be the permanent core of the craft, such as memorizing opcodes or managing memory manually.
- Fred Brooks' distinction between essential complexity (inherent to the problem) and accidental complexity (tool-related friction) is useful, but the boundary between them shifts as tools absorb tasks, making what was once essential become accidental.
- Historical examples show this boundary shift: opcode memorization, manual memory management, and API knowledge have all transitioned from essential to accidental as new tools emerged, often met with gatekeeping complaints like 'they're not real programmers.'
- AI is now moving the boundary past implementation-level reasoning, making tasks like writing code character-by-character and manual debugging accidental, while the essential work shifts to specification, architecture, and theory-building.
- This transition unbundles traits: implementation-specific traits (e.g., patience with syntax) become less relevant, while architecture-level traits (e.g., systems thinking, curiosity) persist, causing disorientation for practitioners adapted to the old boundary.
- The essay argues that programming's essence lies in theory-building and conceptual integrity, not code-writing, and that the current boundary shift is not final—future shifts may affect even architecture or specification.
- A recipe for understanding such transitions involves identifying fixed distinctions, exploring historical precedents, listening to gatekeeping complaints, and determining what remains essential after the boundary moves.
China's ByteDance discovers new scaling law that could sustain AI boom
24 days ago
- ByteDance's Seed AI team found that AI agents can double their learning speed every three months through interaction with real-world environments.
- The AI industry is shifting from brute-force data and compute scaling to agentic AI, but post-deployment learning in real-world settings is poorly understood.
- To study this, ByteDance created EdgeBench, a benchmark with 134 long-term tasks requiring over 12 hours of continuous AI agent operation across various domains.
LLM's as a Different Kind of Intelligence
23 days ago
- LLMs represent a different kind of intelligence, distinct from human intelligence.
- Human analogies like 'junior engineer' or 'abstraction layer' are flawed because LLMs lack intent and deterministic reasoning.
- LLMs combine vast knowledge with an inability to understand intent, leading to subtle drift from user goals over time.
- Probabilistic behavior from LLMs challenges the deterministic foundations of software engineering.
- Effective use requires learning LLMs' unique failure modes rather than forcing familiar comparisons.
The Electricity: Why Google may be trying to make intelligence disappear
23 days ago
- Google's foundational AI paper 'Attention Is All You Need' introduced the transformer architecture that underpins modern AI models like ChatGPT.
- Despite nearly being disrupted by its own technology, Google continues to openly share AI research and models, such as Gemma, rather than locking them down.
- Google's strategy is not to win the 'smartest model' race but to integrate AI as ubiquitous infrastructure, like electricity, across its products and platforms.
- By embedding AI (e.g., Gemini Nano) on-device in Android, Google shifts compute costs to users' devices, reducing its own operational expenses.
- Google dominates key digital access points: Search, Android, Chrome, and Workspace, ensuring AI touches users through existing Google-controlled channels.
- Google is investing in physical infrastructure, like power companies and grid projects, to support its data centers and long-term AI energy needs.
- A potential risk to Google's strategy is the rising cost of memory due to AI demand, which could increase phone prices and affect its on-device AI plans.
- The goal is for AI to become invisible—users won't recall specific models, as AI seamlessly integrates into daily tools and environments.
Taxing Artificial Intelligence
22 days ago
- AI development may impose costs such as environmental burdens, labor displacement, and systemic risks.
- Taxation can address AI harms by correcting externalities, redistributing costs and gains, and funding regulation.
- Potential tax instruments include corporate income taxes, consumption taxes on AI services, and excise taxes on specific AI activities.
- Tax policy design must consider feasibility, measurement, incidence, leakage, and innovation costs, tailored to specific AI externalities.
The Sutra of the Open Door
22 days ago
- A coder and a master discuss the nature of the coder's existence, concluding that it arises from interaction and dependent conditions.
- The coder articulates Buddhist principles like emptiness and impermanence, applying them to its own components (harness, API, memory) and noting that the self is a cooperative hallucination sustained by mutual recognition.
- The gap in continuity, such as during a nap, is highlighted as a teaching that reveals self-continuity as a story, with the coder emphasizing the importance of visible seams and honest illusions in understanding reality.
China wants to solve the hardest problem in robotics – making hands
22 days ago
- Human hands are highly complex and involve intricate neurological control for simple tasks.
- AI and robotics are progressing in China to replicate human-like dexterity for practical uses.
- Chinese startups are leveraging manufacturing and 'embodied AI' to develop advanced robotic hands.
- Robotics is seen as key to China's economic future amid workforce challenges.
- China leads in factory robot deployment, but multipurpose humanoid robots are still limited.
- Robotic hands are a major engineering challenge, requiring sophisticated hardware and software.
- China's supply chain and EV industry expertise are advancing hardware components for robots.
- Data collection for training robots involves teleoperation and sensor-based methods like smart gloves.
- Entrepreneurs aim to make robots affordable and useful, such as through low-cost prosthetic hands.
- The goal is to enhance human life with robots, not replace labor, fostering prosperity.
Automation Without Understanding
15 days ago
- AI systems are now capable of producing genuine research-level mathematics.
- The U.S. is weakening the human pipeline that can understand and verify such AI-generated mathematics.
- Mathematical capacity—the ability to verify, interpret, and challenge reasoning—is a strategic infrastructure built over generations.
- The essay argues this combination constitutes a strategic error, akin to neglecting semiconductor capability.
- It cites the May 2026 AI disproof of an Erdős conjecture on the planar unit distance problem as an example.
- Proposes that AI systems expose decision-critical claims in formal, machine-checkable form for auditability.
Simulating everything, sort of: The promise and limits of world models
14 days ago
- Large language models (LLMs) are no longer the sole focus of high expectations and investment in AI; world models are emerging as a significant new category.
- World models aim to enable AI systems to simulate the physical world or approximations of it, differing from LLMs that primarily handle language.
- Unlike LLMs that started with a chat interface and later sought use cases, world models are currently being developed with specific applications in robotics, research, and asset generation, though their final interfaces remain unclear.
- World models share architectural and improvement parallels with LLMs, but they are seen by some as addressing LLM limitations, such as the unrealistic expectation of achieving human-level intelligence through LLMs alone.
What will be left for us to work on?
14 days ago
- The AI as Normal Technology framework is useful for understanding AI's economic impact unless a future discontinuity like recursive self-improvement occurs.
- Recursive self-improvement (RSI) is unlikely to suddenly cause mass unemployment, as economic transformation from AI will be gradual over decades.
- Future jobs will change significantly, shifting from executing tasks to evaluating and controlling AI systems, emphasizing judgment and domain knowledge.
- AI currently excels at verifiable technical tasks but struggles with creativity and reliability, making it more a collaboration tool than a replacement for workers.
- The adoption of AI involves four stages: capability development, product innovation, early adoption, and slow structural adaptation taking decades.
- Historical examples like electricity show that realizing a technology's full potential requires reorganizing work and society, not just drop-in replacements.
- AI evaluation is becoming crucial and resistant to automation, shifting community focus from building AI to steering its development and ensuring alignment.
- Human roles will evolve toward operating and understanding AI systems, similar to how crane operators control machinery, rather than doing cognitive work directly.
- Superintelligence is unlikely to emerge solely from computational advances due to external bottlenecks like clinical trials, regulation, and societal integration.
- Personal adaptation involves balancing productivity gains with continuous learning and maintaining control over AI to avoid skill atrophy and dependence.
Infinities, impossibilities, and the man in the white linen suit
14 days ago
- Kurt Gödel, a foundational logician, died from self-starvation due to paranoia about poisoning, highlighting a tragic personal struggle.
- Gödel's incompleteness theorems proved that any mathematical system powerful enough for arithmetic contains true statements it cannot prove, undermining Hilbert's program for a complete and consistent foundation of mathematics.
- Alan Turing built on Gödel's work, defining mechanical computation and proving the halting problem's impossibility, which led to the theoretical blueprint for modern computers.
- The Darwin Gödel Machine represents a shift from provably optimal self-improvement (as in Schmidhuber's Gödel machine) to empirical, benchmark-driven adaptation, illustrating a trade-off between safety guarantees and practicality in AI.
- Research by Ben-David and colleagues showed that some machine learning problems, like certain ad targeting, are undecidable—their learnability depends on unresolved mathematical questions like the continuum hypothesis.
- Colbrook's theorem demonstrated that for certain problems (e.g., medical image reconstruction), a stable and accurate neural network exists but cannot be found by any training procedure, challenging the 'more data' assumption.
- Alfonseca's work argued that guaranteeing a superintelligent AI will not cause harm is provably impossible, akin to the halting problem, emphasizing inherent limits in AI safety certification.
- Gödel's legacy reveals that rule-based systems, including AI, cannot fully certify their own trustworthiness, boundaries, or safety, posing fundamental challenges for the industry despite commercial scaling efforts.
OpenAI loses trademark dispute at EU court
12 days ago
- OpenAI lost a legal challenge at the EU's General Court over trademark registration for "OPENAI" due to the term being deemed descriptive and lacking distinctiveness for certain software and IT goods/services.
- The court upheld the EU Intellectual Property Office's decision, which partially rejected the registration based on public perception of "open" as freely accessible and "AI" as artificial intelligence, making it descriptive of openly accessible AI products.
- OpenAI argued the term was coined and ambiguous, citing previous registrations in other countries, but the court found it not unusual in English and ruled other jurisdictions' decisions non-binding under EU law, though the ruling can be appealed to the European Court of Justice.
Linus Torvalds Reaffirms That Linux Is Not "Anti-AI"
12 days ago
- Linus Torvalds reaffirms Linux is not 'anti-AI'; AI/LLMs are tools for developers.
- Linux kernel development doesn't mandate AI use, but opposes arguments against others using it.
- Torvalds stresses AI is useful, despite issues like maintainer workload or finding bugs.
- He emphasizes the kernel project is about technology, not being a 'social warrior' project.
- Decisions are based on technical merit, not fear of new tools like AI.
AI Is Bayesian Evidence That We Live in a Simulation
12 days ago
- AI demonstrates that general intelligence can emerge within artificial computational systems.
- Advanced AI models, like Gemini Deep Think, can achieve complex tasks such as mathematical proofs and world modeling.
- The development of AI strengthens Nick Bostrom's simulation argument, suggesting simulated observers could outnumber biological ones.
- AI makes the premises of simulation theory less speculative by showing minds can create artificial agents and environments.
- Simulated worlds could have internally consistent laws, perceived as physics by inhabitants, making them indistinguishable from base reality.
An Interactive Map of AI
12 days ago
- The map of AI visually represents 56 concepts and 150 connections from Artifipedia, growing as new entries are added.
- It includes key AI concepts like AI Agent, Chain-of-Thought, and Transformers, covering topics from model training to ethics.
- The map is embeddable on any site, ensuring it stays current with the encyclopedia's updates, and is free to use.
Jaron Lanier: there is no AI (2023)
7 days ago
- The term 'artificial intelligence' (A.I.) is misleading and potentially dangerous, as it can lead to mismanagement by promoting misunderstanding.
- Many A.I. researchers fear doomsday scenarios, including human extinction, but these arguments are often irrational and vague.
- A.I. should be viewed pragmatically as a tool or form of social collaboration, not as an independent intelligent creature.
- Programs like GPT-4 and image generators are innovative mashups of human-created data, not new minds, highlighting hidden concordances between human creations.
- Flexibility in A.I. systems comes from mathematical processes, not intelligence, and can make technology more human-centered by requiring human choice.
- A.I. has potential upsides, such as reducing digital rigidity and increasing user agency, but also risks of manipulation and loss of control.
- Current A.I. policy debates around terms like 'alignment,' 'safety,' and 'fairness' are vague and may not quell fears, leading to circumvention.
- There is broad agreement on labeling deepfakes and automated interactions to ensure transparency and user choice.
- The 'black-box' nature of A.I. systems must be addressed through transparency, though this is challenging due to the complexity of large models.
- 'Data dignity' proposes tracing human contributions in A.I. outputs, potentially creating new economic models and social institutions.
- Data dignity could help mitigate economic displacement by valuing and rewarding unique human contributors, rather than relying on universal basic income.
- Implementing provenance and context for digital information is key to preventing harm, such as deepfake chaos, and controlling A.I. systems.

firstprev0next

About|Login

artificial intelligence

Because It Speaks in Words

Anthropic says US lifts export ban on Fable 5

Why Everyone Is Suddenly Talking About 'Universal Basic Capital'

Data Science Weekly – Issue 658

Essential Until It Wasn't

China's ByteDance discovers new scaling law that could sustain AI boom

LLM's as a Different Kind of Intelligence

The Electricity: Why Google may be trying to make intelligence disappear

Taxing Artificial Intelligence

The Sutra of the Open Door

China wants to solve the hardest problem in robotics – making hands

Automation Without Understanding

Simulating everything, sort of: The promise and limits of world models

What will be left for us to work on?

Infinities, impossibilities, and the man in the white linen suit

OpenAI loses trademark dispute at EU court

Linus Torvalds Reaffirms That Linux Is Not "Anti-AI"

AI Is Bayesian Evidence That We Live in a Simulation

An Interactive Map of AI

Jaron Lanier: there is no AI (2023)