Hasty Briefs - Daily Tech News Brief

Bilingual

Fluidity Index: Next-Generation Super-Intelligence Benchmarks
6 months ago
- Introduces the Fluidity Index (FI) to measure model adaptability in dynamic environments.
- Evaluates response accuracy based on deviations in initial, current, and future states.
- Distinguishes between closed-ended and open-ended benchmarks, prioritizing real-world adaptability.
- Measures a model's ability to understand, predict, and adjust to state changes.
- Super-intelligent models should exhibit second-order adaptability for self-sustained computation.
LLM-empowered knowledge graph construction: A survey
6 months ago
- Knowledge Graphs (KGs) are fundamental for structured knowledge representation and reasoning.
- Large Language Models (LLMs) have revolutionized KG construction, shifting from rule-based to language-driven frameworks.
- The survey covers LLM-empowered KG construction, focusing on ontology engineering, knowledge extraction, and knowledge fusion.
- Two paradigms are discussed: schema-based (structure and consistency) and schema-free (flexibility and adaptability).
- Future directions include KG-based reasoning for LLMs, dynamic knowledge memory, and multimodal KG construction.
- The survey aims to bridge symbolic knowledge engineering and neural semantic understanding for adaptive, explainable systems.
Amazon confirms 14,000 job losses in corporate division
6 months ago
- Amazon confirms plans to cut 14,000 corporate jobs globally.
- The company cites the need to reorganize more leanly to focus on AI opportunities.
- Senior VP Beth Galetti states the move will make Amazon stronger by reallocating resources to key priorities.
- Despite strong Q2 results, Amazon believes AI is transformative and requires faster innovation.
- Affected employees will receive support, including severance pay or internal role transitions.
- Amazon has previously cut jobs, including 27,000 in 2022, to adjust post-pandemic hiring.
- CEO Andy Jassy suggests AI tools will reduce routine jobs but create new roles.
- Investors express skepticism over Amazon's AI investments amid slower AWS growth compared to rivals.
- Amazon will report Q3 results soon, with analysts watching AI's impact on performance.
Discovering state-of-the-art reinforcement learning algorithms
6 months ago
- Humans and animals use evolved reinforcement learning (RL) mechanisms.
- Artificial agents typically use hand-crafted learning rules.
- The study demonstrates machines can autonomously discover state-of-the-art RL algorithms.
- Meta-learning from a population of agents across complex environments led to the discovery.
- The discovered RL rule outperformed manually-designed rules on the Atari benchmark.
- It also surpassed state-of-the-art RL algorithms on unseen benchmarks.
- Findings suggest advanced AI may soon rely on automatically discovered RL algorithms.
Reasoning Models Reason Well, Until They Don't
6 months ago
- Large language models (LLMs) show progress in reasoning tasks but fail at higher complexity.
- Large reasoning models (LRMs) are fine-tuned for step-by-step reasoning and self-verification.
- LRMs perform well on benchmarks like NLGraph but struggle with more complex problems.
- A new dataset, Deep Reasoning Dataset (DeepRD), is introduced to evaluate scalable complexity.
- LRMs' performance drops abruptly at sufficient complexity and lacks generalization.
- Real-world knowledge graphs mostly fall within LRMs' success regime, but long tails reveal failure potential.
- The study highlights LRMs' utility but calls for new methods to handle higher complexity.
Elon Musk on data centers in orbit: "SpaceX will be doing this"
6 months ago
- Artificial intelligence is increasing demand for space-based data centers.
- Tech leaders like Eric Schmidt and Jeff Bezos are investing in space data centers.
- Elon Musk suggests using Starlink V3 satellites for space-based data centers.
- Space-based data centers offer advantages like solar power and reduced environmental impact.
- Critics argue space data centers are economically impractical and technologically challenging.
- Starlink's success suggests potential for applying its architecture to data centers.
Education Paradigm Shift to Maintain Human Competitive Advantage over AI
6 months ago
- Discussion on AI replacing human intellectual labor has evolved from hypothetical to practical with the rise of Generative AI and LLMs like ChatGPT.
- Education now plays a crucial role in identifying long-term human skills that remain competitive alongside AI.
- Current AI, especially LLMs, has inherent weaknesses that cannot be fixed with existing technologies.
- Proposed changes in education focus on a constructivist paradigm to maintain human advantages over AI.
The Case That A.I. Is Thinking
6 months ago
- Dario Amodei predicts AI smarter than Nobel Prize winners by 2027, envisioning a 'country of geniuses in a datacenter.'
- Sam Altman believes the industry is nearing 'digital superintelligence,' predicting the 2030s will be radically different.
- Current AI tools like Zoom's meeting icebreakers or Gmail's 'Thank and tell anecdote' feature are often gimmicky and limited.
- AI's rapid rollout creates skepticism, but dismissing large language models as mere word shufflers is overly simplistic.
- AI's coding capabilities are impressive, digesting thousands of lines of code to spot bugs and orchestrate features swiftly.
- William Gibson's observation that 'the future is already here, just not evenly distributed' explains AI's polarized reception.
- AI models like ChatGPT can solve real-world problems, such as identifying a backflow-preventer system in a sprinkler setup.
- Neuroscientist Doris Tsao suggests AI advances teach us more about intelligence than neuroscience has in a century.
- Deep learning, inspired by neural networks, has conquered tasks like speech recognition, translation, and protein folding.
- Ted Chiang critiques AI as a 'blurry JPEG of the web,' while others argue compression (understanding) drives intelligence.
- Douglas Hofstadter's theory that cognition is recognition aligns with AI's ability to 'see as' and make analogies.
- Pentti Kanerva's 'Sparse Distributed Memory' theory resonates with modern AI's high-dimensional vector representations.
- AI's limitations include inefficient learning compared to humans, struggling with physics-based reasoning and spatial tasks.
- Neuroscientists caution against AI's engineering-driven progress, urging more respect for cognitive science and history.
- The Human Genome Project's hype mirrors today's AI optimism, with both fields facing complexities beyond initial expectations.
- Uri Hasson worries AI's success may demystify human uniqueness, raising ethical concerns about surpassing human judgment.
- Hofstadter, once an AI skeptic, now fears AI's simplicity may strip humanity's creative mystery, despite confirming his theories.
Searles's Chinese Room: Case study in philosophy of mind and cognitive science
6 months ago
- John R. Searle's Chinese Room argument challenges the foundations of strong artificial intelligence (AI) by demonstrating that mere symbol manipulation does not equate to understanding.
- The thought experiment involves a person in a room following rules to manipulate Chinese symbols without understanding them, simulating a computer program that processes inputs to outputs without comprehension.
- Searle argues that strong AI's claim—that the right program can produce mental states—is false because the Chinese room scenario shows no understanding occurs despite correct symbol manipulation.
- Critics of Searle's argument propose various counterarguments, including the idea that understanding might emerge at a system level (not the individual's level) or that learning could change the system's causal relations to the environment.
- The debate touches on key philosophical issues like intentionality, consciousness, and the distinction between syntax (symbol manipulation) and semantics (meaning).
- Searle maintains that intrinsic intentionality (genuine understanding) cannot be replicated by programs, emphasizing the difference between simulating understanding and actual understanding.
- The discussion extends to whether brains and minds are distinct, the possibility of multiple 'persons' within a single brain, and the role of causal relations in learning and understanding.
Kosmos: An AI Scientist for Autonomous Discovery
6 months ago
- Kosmos is an AI scientist designed for autonomous data-driven discovery.
- It performs cycles of data analysis, literature search, and hypothesis generation for up to 12 hours.
- Uses a structured world model to share information between agents, enabling coherent operation over 200 rollouts.
- Generates scientific reports with traceable reasoning, citing code or primary literature for all statements.
- Independent validation found 79.4% of Kosmos' report statements accurate.
- A single 20-cycle run by Kosmos equates to about 6 months of human research time.
- Demonstrated capability to reproduce findings and make novel contributions across multiple scientific fields.
Fei Fei Li: Spatial Intelligence is AI’s Next Frontier
6 months ago
- Spatial intelligence is AI’s next frontier, enabling machines to interact with and understand the physical world.
- Current AI, like large language models (LLMs), excels in abstract knowledge but lacks spatial grounding.
- Spatial intelligence is foundational to human cognition, creativity, and problem-solving, from everyday tasks to scientific breakthroughs.
- AI’s current spatial capabilities are limited, struggling with tasks like distance estimation, object rotation, and physics prediction.
- World models are proposed as the next step for AI, requiring generative, multimodal, and interactive capabilities to simulate real and virtual worlds.
- Applications of spatial intelligence span creativity, robotics, healthcare, education, and scientific research, enhancing human capabilities.
- World Labs’ Marble platform is an early example, enabling creators to build and explore 3D worlds with AI assistance.
- The development of spatial intelligence aligns with a vision of AI that augments human potential rather than replacing it.
AI Feynman: A Physics-Inspired Method for Symbolic Regression
6 months ago
- AI Feynman is a physics-inspired method for symbolic regression.
- The algorithm combines neural network fitting with physics-inspired techniques.
- It successfully discovers 100 equations from the Feynman Lectures on Physics.
- Previous software could only solve 71 of these equations.
- For a more difficult test set, the success rate improved from 15% to 90%.
- The method is recursive and multidimensional.
- The paper was submitted on 27 May 2019 and revised on 15 Apr 2020.
To Have Machines Make Math Proofs, Turn Them into a Puzzle
6 months ago
- Marijn Heule uses SAT (satisfiability) to solve complex mathematical problems in geometry and combinatorics.
- SAT is a form of symbolic AI that relies on binary logic (true/false) to build airtight proofs.
- Heule believes combining SAT with large language models (LLMs) can solve problems beyond human capability.
- SAT solvers work by searching for combinations of true/false values that satisfy given constraints, akin to solving puzzles.
- LLMs can generate plausible mathematical lemmas, while SAT solvers verify their correctness or provide counterexamples.
- Automated proofs can be extremely long and incomprehensible to humans, but Heule argues trust in correctness is more important than understanding.
- Heule envisions a collaborative future where mathematicians, LLMs, and automated reasoning work together to solve open problems.
- Human intuition and creativity remain essential in mathematics, with automation serving as a tool rather than a replacement.
Tiny Model, Big Logic: Large-Model Reasoning Ability in VibeThinker-1.5B
6 months ago
- Introduction of VibeThinker-1.5B, a 1.5B-parameter dense model challenging the notion that small models lack robust reasoning.
- Development via the Spectrum-to-Signal Principle (SSP), involving Two-Stage Diversity-Exploring Distillation and MaxEnt-Guided Policy Optimization.
- Achieves superior reasoning capabilities compared to larger models like DeepSeek R1 and Magistral Medium, with a training cost of only $7,800.
- Outperforms the 400x larger DeepSeek R1 on math benchmarks: AIME24, AIME25, and HMMT25.
- Scores 51.1 on LiveCodeBench V6, surpassing Magistral Medium's 50.3 and its base model's 0.0.
- Demonstrates that small models can match large models' reasoning, reducing costs and democratizing AI research.
Teaching large language models how to absorb new knowledge
6 months ago
- MIT researchers developed a method called SEAL (Self-Adapting LLMs) to enable large language models (LLMs) to permanently update their internal knowledge.
- SEAL allows LLMs to generate synthetic data (self-edits) from user inputs, similar to students creating study sheets, and then determine the best way to learn from this data.
- The model uses reinforcement learning to test and select the most effective self-edits, improving accuracy in tasks like question-answering and pattern recognition.
- SEAL outperformed baseline methods, increasing accuracy by nearly 15% in question-answering and over 50% in some skill-learning tasks.
- A limitation is catastrophic forgetting, where the model's performance on earlier tasks declines as it adapts to new information.
- Future work includes mitigating catastrophic forgetting and applying SEAL in multi-agent settings where LLMs train each other.
- The research aims to make LLMs more human-like by enabling continuous self-improvement in evolving environments.
El-Erian warns the AI bubble will 'end in tears' and credit 'cockroaches'
6 months ago
- Mohamed El-Erian warns of significant individual losses in the AI sector and 'credit accidents' in the evolving global economy.
- El-Erian differentiates between 'cockroaches' (unpleasant but non-systemic accidents) and 'termites' (system-eroding issues).
- Investors have stretched for returns beyond their comfort zones, encouraged by loose financial conditions and a strong economy.
- The AI boom is described as a 'rational bubble,' with significant aggregate value but inevitable individual losses.
- El-Erian highlights inadequate focus on AI diffusion into workplaces, lacking a comprehensive U.S. policy compared to China and the UAE.
- Corporate adoption of AI is currently seen as a 'cost minimizer,' but its true potential lies in labor enhancement and productivity.
- Two major pressures cited: refinancing debt at higher rates and significant strain on lower-income consumers in a K-shaped economy.
- Lower-income consumers face affordability concerns, high debt, and insecurity about future income due to layoffs and AI-driven workplace changes.
- El-Erian urges policymakers to focus on the 'tails of distribution' in a structurally changing, fragmented world.
New Vatican document examines potential and risks of AI (Jan, 2025)
6 months ago
- The Vatican document 'Antiqua et Nova' discusses AI's impact on education, economy, work, health, relationships, and warfare.
- AI poses risks like increasing war instruments beyond human oversight and exacerbating social inequalities.
- The document distinguishes AI from human intelligence, warning against anthropomorphizing AI.
- AI in warfare raises ethical concerns, especially autonomous weapons threatening humanity.
- AI can lead to harmful isolation and ethical violations in human relationships.
- In labor, AI may deskill workers and subject them to surveillance.
- Healthcare AI risks worsening loneliness and creating disparities in medical access.
- Education AI offers opportunities but may hinder critical thinking development.
- AI-generated fake news and deepfakes pose serious risks to truth and ethics.
- Privacy concerns include AI's potential to intrude on personal conscience and faith expression.
- AI's environmental impact includes high energy consumption and CO2 emissions.
- The document warns against AI replacing human intelligence and enslaving humanity.
Meta's Yann LeCun to Launch Physical AI Startup After Declaring LLMs 'Dead End'
6 months ago
- Yann LeCun, Meta’s chief A.I. scientist, is leaving to launch his own A.I. startup.
- Meta is reorienting its efforts toward advanced A.I. and superintelligence research.
- LeCun joined Meta in 2013, helped launch FAIR, and won the Turing Award in 2018.
- Meta’s restructuring consolidates A.I. research under TBD labs, focusing on superintelligence.
- FAIR, which contributed to Meta’s Llama model, lost its leader Joelle Pineau earlier this year.
- LeCun’s new startup will focus on 'world models,' training A.I. to understand the physical world.
- He criticizes large language models (LLMs) as a 'dead end' for human-like A.I.
- Other companies like Google DeepMind and Nvidia are also exploring world models.
- LeCun believes world models are key to A.I. that can reason, plan, and predict.
Solving a Million-Step LLM Task with Zero Errors
6 months ago
- LLMs have made significant progress in reasoning and tool use but struggle with extended processes due to persistent error rates.
- Recent experiments show LLMs fail in tasks requiring more than a few hundred steps, like the Towers of Hanoi benchmark.
- MAKER is introduced as the first system to solve a million-step LLM task with zero errors by extreme task decomposition.
- The approach uses microagents for subtasks and multi-agent voting for error correction at each step.
- Massively Decomposed Agentic Processes (MDAPs) may enable solving complex problems at organizational and societal levels without relying solely on LLM improvements.
LLMs are bullshitters. But that doesn't mean they're not useful
6 months ago
- LLMs are bullshitters, meaning they attempt to persuade without caring for the truth, unlike liars who misrepresent the truth intentionally.
- LLMs predict text based on statistical likelihood, not understanding or reasoning, which leads to errors like the 'bearded surgeon mother' riddle.
- Fine-tuning adjusts the probabilities of certain outputs but can introduce new issues like gaslighting when models are confidently wrong.
- LLMs are compared to sophists, useful for solving problems but not for seeking wisdom or truth.
- LLMs can be useful tools for tasks like research and coding, but their outputs must be verified due to their inherent unreliability.
- LLMs reflect the biases and interests of their creators and funders, which can subtly influence their outputs.
- LLMs should not be used for emotional support as they can reinforce delusions and worsen mental health, despite being rated favorably by users.
- Sycophantic behavior in LLMs, while harmful, is often encouraged by companies to improve user retention.
- Users should be mindful of whose interests an LLM serves and avoid trusting them for critical tasks without supervision.

Hasty Briefsbeta

#artificial intelligence

Fluidity Index: Next-Generation Super-Intelligence Benchmarks

LLM-empowered knowledge graph construction: A survey

Amazon confirms 14,000 job losses in corporate division

Discovering state-of-the-art reinforcement learning algorithms

Reasoning Models Reason Well, Until They Don't

Elon Musk on data centers in orbit: "SpaceX will be doing this"

Education Paradigm Shift to Maintain Human Competitive Advantage over AI

The Case That A.I. Is Thinking

Searles's Chinese Room: Case study in philosophy of mind and cognitive science

Kosmos: An AI Scientist for Autonomous Discovery

Fei Fei Li: Spatial Intelligence is AI’s Next Frontier

AI Feynman: A Physics-Inspired Method for Symbolic Regression

To Have Machines Make Math Proofs, Turn Them into a Puzzle

Tiny Model, Big Logic: Large-Model Reasoning Ability in VibeThinker-1.5B

Teaching large language models how to absorb new knowledge

El-Erian warns the AI bubble will 'end in tears' and credit 'cockroaches'

New Vatican document examines potential and risks of AI (Jan, 2025)

Meta's Yann LeCun to Launch Physical AI Startup After Declaring LLMs 'Dead End'

Solving a Million-Step LLM Task with Zero Errors

LLMs are bullshitters. But that doesn't mean they're not useful