Import AI

Import AI is a comprehensive newsletter that covers advancements and discussions in artificial intelligence, focusing on emerging research, AI policy changes, technological developments, and the ethical implications of AI. It provides analysis on AI models, safety risks, AI applications in various domains, and global perspectives on AI's future.

AI Models and Frameworks AI Policy and Safety Technological Developments in AI Ethical and Societal Implications of AI Large Language Models Artificial General Intelligence AI in Scientific Research Global Perspectives on AI

The hottest Substack posts of Import AI

And their main takeaways

Import AI 374: China's military AI dataset; platonic AI; brainlike convnets

339 implied HN points • 27 May 24
🕹 Technology AI Research Search Engines Neural Networks Tech Policy
  1. UC Berkeley researchers discovered a suspicious Chinese military dataset named 'Zhousidun' with specific images of American destroyers, presenting potential implications for military use of AI.
  2. Research suggests that as AI systems scale up, their representations of reality become more similar, with bigger models better approximating the world we exist in.
  3. Convolutional neural networks are shown to align more with primate visual cortexes than transformers, indicating architectural biases that can lead to better understanding the brain.

Import AI 373: Guaranteed safety; West VS East AI attitudes; MMLU-Pro

419 implied HN points • 20 May 24
🕹 Technology AI Research Benchmarking Regulation
  1. Academic researchers have built the National Deep Inference Fabric (NDIF) to experiment with large-scale AI models in a transparent manner.
  2. Researchers have outlined a framework for building 'guaranteed safe' AI systems, involving components like safety specifications, world models, and verifiers.
  3. A global survey indicates that Western countries have more pessimism towards AI regulation compared to China and India, potentially changing how governments approach regulating and adopting AI.

Import AI 372: Gibberish jailbreak; DeepSeek's great new model; Google's soccer-playing robots

399 implied HN points • 13 May 24
🕹 Technology AI Research Language Models Deep Learning Simulation Ethics
  1. DeepSeek released a powerful language model called DeepSeek-V2 that surpasses other models in efficiency and performance.
  2. Research from Tsinghua University shows how mixing real and synthetic data in simulations can improve AI performance in real-world tasks like medical diagnosis.
  3. Google DeepMind trained robots to play soccer using reinforcement learning in simulation, showcasing advancements in AI and robotics;

Import AI 371: CCP vs Finetuning; why people are skeptical of AI policy; a synthesizer for a LLM

439 implied HN points • 06 May 24
🕹 Technology AI Research Data Analysis Medical AI Image Generation Internet culture
  1. People are skeptical of AI safety policy as different views arise from the same technical information, making it important to consider varied perspectives.
  2. Chinese researchers have developed a method called SOPHON to openly release AI models while preventing finetuning for misuse, offering a solution for protecting against subsequent harm.
  3. Automating intelligence analysis through datasets like OpenStreetView-5M will enhance training machine learning systems for geolocation, leading to potential applications in both military intelligence and civilian sectors.

Import AI 370: 213 AI safety challenges; everything becomes a game; Tesla's big cluster

439 implied HN points • 29 Apr 24
🕹 Technology AI Data Security Research Systems
  1. Chinese researchers introduced MMT-Bench, a benchmark for assessing visual reasoning in language models with diverse tasks and scenarios.
  2. Researchers developed a system to turn 2D photos into 3D gameworlds, showing AI's capability to transform real-world imagery into interactive experiences.
  3. A consortium of researchers addressed 213 AI safety challenges across 18 areas, emphasizing the urgent need for solutions to ensure the reliability and safety of language models.
Get a weekly roundup of the best Substack posts, by hacker news affinity:

Import AI 369: Conscious machines are possible; AI agents; the varied uses of synthetic data

539 implied HN points • 15 Apr 24
🕹 Technology AI
  1. Synthetic data is crucial in AI development, allowing for the generation of additional data without relying solely on human input.
  2. OSWorld showcases how AI systems can potentially become integrated into daily computer tasks, creating a future where AI is ever-present in our interactions with technology.
  3. Research suggests that the development of conscious machines may be feasible, exploring theories on machine consciousness and potential capabilities.

Import AI 357: Facebook's open source AGI plan; Google beats humans at geometry problems; and Intel makes its GPUs better

2076 implied HN points • 22 Jan 24
🕹 Technology AI Research Machine Learning Machine Translation
  1. Facebook aims to develop artificial general intelligence (AGI) and make it open-source, marking a significant shift in focus and possibly accelerating AGI development.
  2. Google's AlphaGeometry, an AI for solving geometry problems, demonstrates the power of combining traditional symbolic engines with language models to achieve algorithmic mastery and creativity.
  3. Intel is enhancing its GPUs for large language models, a necessary step towards creating a competitive GPU offering compared to NVIDIA, although the benchmarks provided are not directly comparable to industry standards.

Import AI 368: 500% faster local LLMs; 38X more efficient red teaming; AI21's Frankenmodel

559 implied HN points • 08 Apr 24
🕹 Technology AI Research AI Models AI Policy Robotics Artificial Intelligence
  1. Efficiency improvements can be achieved in AI systems by varying the frequency at which GPUs operate, especially for tasks with different input and output lengths.
  2. Governments like Canada are investing significantly in AI infrastructure and safety measures, reflecting the growing importance of AI in economic growth and policymaking.
  3. Advancements in AI technologies are making it easier for individuals to run large language models locally on their own machines, leading to a more decentralized access to AI capabilities.

Import AI 367: Google's world-spanning model; breaking AI policy with evolution; $250k for alignment benchmarks

599 implied HN points • 01 Apr 24
🕹 Technology AI Supercomputers Benchmarks Cybersecurity
  1. Google is working on a distributed training approach named DiPaCo to create large neural networks that break traditional AI policy focusing on centralized models.
  2. Microsoft and OpenAI plan to build a $100 billion supercomputer for AI training, signaling the transition of AI industry towards capital intensive endeavors like oil extraction or heavy industry, touching on regulatory and industrial policy implications.
  3. Sakana AI has developed 'Evolutionary Model Merge' method to create advanced AI models by combining existing ones through evolutionary techniques, potentially changing AI policy by challenging the need for costly model development.

Import AI 356: China's good LLM; AI credit scores; and fooling VLMs with REBUS

1238 implied HN points • 15 Jan 24
🕹 Technology AI Research Language Models Compute Robotics
  1. Today's AI systems struggle with word-image puzzles like REBUS, highlighting issues with abstraction and generalization.
  2. Chinese researchers have developed high-performing language models similar to GPT-4, showing advancements in the field, especially in Chinese language processing.
  3. Language models like GPT-3.5 and 4 can already automate writing biological protocols, hinting at the potential for AI systems to accelerate scientific experimentation.

Import AI 364: Robot scaling laws; human-level LLM forecasting; and Claude 3

519 implied HN points • 11 Mar 24
🕹 Technology AI Robots Generative models Biology
  1. Scaling laws are transforming the world of robotics - more data, bigger context windows, and more parameters in models lead to significant improvements quickly.
  2. Advancements in AI forecasting show that language models can match human capabilities in predicting binary outcomes, suggesting a future of accurate forecasting by AI systems.
  3. New datasets like Panda-70M for video captioning and models like Evo for biological predictions are pushing the boundaries of AI and demonstrating the power of generative models in various domains.

Import AI 354: Distributed LLM inference; CCP-approved dataset; AI scientists

1278 implied HN points • 25 Dec 23
🕹 Technology AI Research Language Models Datasets
  1. Distributed inference is becoming easier with AI collectives, allowing small groups to work with large language models more efficiently and effectively.
  2. Automation in scientific experimentation is advancing with large language models like Coscientist, showcasing the potential for LLMs to automate parts of the scientific process.
  3. Chinese government's creation of a CCP-approved dataset for training large language models reflects the move towards LLMs aligned with politically correct ideologies, showcasing a unique approach to LLM training.

Import 355: Local LLMs; scaling laws for inference; free Mickey Mouse

1058 implied HN points • 08 Jan 24
🕹 Technology AI Robotics Programming AI Research Computing
  1. PowerInfer software allows $2k machines to perform at 82% of the performance of $20k machines, making it more economically sensible to sample from LLMs using consumer-grade GPUs.
  2. Surveys show that a significant number of AI researchers worry about extreme scenarios such as human extinction from advanced AI, indicating a greater level of concern and confusion in the AI development community than popular discourse suggests.
  3. Robots are becoming cheaper for research, like Mobile ALOHA that costs $32k, and with effective imitation learning, they can autonomously complete tasks, potentially leading to more robust robots in 2024.

Import AI 365: WMD benchmark; Amazon sees $1bn training runs; DeepMind gets closer to its game-playing dream

399 implied HN points • 18 Mar 24
🕹 Technology AI Research Robotics AI training
  1. Alliance for the Future (AFTF) was founded in response to concerns about overreach in AI safety regulation, highlighting the importance of well-intentioned policies leading to counter-reactions.
  2. Covariant's RFM-1 shows how generative AI can be applied to industrial robots, allowing easy robot operation through human-like instructions, reflecting a shift towards faster-moving robotics facilitated by AI.
  3. DeepMind's SIMA represents a significant advancement towards a general AI agent by fusing recent AI advancements, showcasing the potential of scaling up diverse AI functions in new environments, opening possibilities for further development and complexity.

Import AI 363: ByteDance's 10k GPU training run; PPO vs REINFORCE; and generative everything

419 implied HN points • 04 Mar 24
🕹 Technology AI Research Reinforcement Learning Language Models Ethics
  1. DeepMind developed Genie, a system that transforms photos or sketches into playable video games by inferring in-game dynamics.
  2. Researchers found that for language models, the REINFORCE algorithm can outperform the widely used PPO, showing the benefit of simplifying complex processes.
  3. ByteDance conducted one of the largest GPU training runs documented, showcasing significant non-American players in large-scale AI research.

Import AI 361: GPT-4 hacking; theory of minds in LLMs; and scaling MoEs + RL

359 implied HN points • 19 Feb 24
🕹 Technology AI Research Cybersecurity Multimodal models Language Models
  1. Researchers have discovered how to scale up Reinforcement Learning (RL) using Mixture-of-Experts models, potentially allowing RL agents to learn more complex behaviors.
  2. Recent research shows that advanced language models like GPT-4 are capable of autonomous hacking, raising concerns about cybersecurity threats posed by AI.
  3. Adapting off-the-shelf AI models for different tasks, even with limited computational resources, is becoming easier, indicating a proliferation of AI capabilities for various applications.

Import AI 360: Guessing emotions; drone targeting dataset; frameworks for AI alignment

379 implied HN points • 12 Feb 24
🕹 Technology AI Datasets Security Language Models Ethics
  1. Teaching AI to understand complex human emotions like joy, surprise, and anger can help in applications like surveillance and advertising.
  2. AI systems, like other software, are vulnerable to attacks, as shown by a demonstration breaking MoE models with a buffer overflow attack.
  3. Frameworks are being developed to ensure AI systems align with diverse human values, considering various perspectives and how to measure alignment.
  4. The development of AI systems is advancing in areas like emotion recognition, system security, and value alignment.
  5. Researchers are pushing the boundaries of AI capabilities, from emotion recognition to security to ethical alignment.
  6. Current AI trends indicate growth in researching human emotions, security vulnerabilities, and ethical considerations.

Import AI 362: Amazon's big speech model; fractal hyperparameters; and Google's open models

299 implied HN points • 26 Feb 24
🕹 Technology AI Models Language Models Fiction
  1. The full capabilities of today's AI systems are still not fully explored, with emerging abilities seen as models scale up.
  2. Google released Gemma, small but powerful AI models that are openly accessible, contributing to the competitive AI landscape.
  3. Understanding hyperparameter settings in neural networks is crucial as the fine boundary between stable and unstable training is found to be fractal, impacting the efficiency of training runs.

Import AI 359: $1 billion gov supercomputer; Apple’s good synthetic data technique; and a thousand-year old data library

339 implied HN points • 05 Feb 24
🕹 Technology AI Research Supercomputers Data Storage Language Models
  1. Google uses LLM-powered bug fixing that is more efficient than human fixes, highlighting the impact of AI integration in speeding up processes.
  2. Yoshua Bengio suggests governments invest in supercomputers for AI development to stay ahead in monitoring tech giants, emphasizing the importance of AI investment in the public sector.
  3. Microsoft's Project Silica showcases a long-term storage solution using glass for archiving data, which is a unique and durable alternative to traditional methods.
  4. Apple's WRAP technique creates synthetic data effectively by rephrasing web articles, enhancing model performance and showcasing the value of incorporating synthetic data in training.

Import AI 353: AI bootstrapping; LLMs as inventors; Facebook releases a free moderation tool

559 implied HN points • 18 Dec 23
🕹 Technology AI Machine Learning Moderation Research Language Models
  1. AI bootstrapping is advancing, with techniques like ReST^EM by Google DeepMind showing ways to make models smarter iteratively.
  2. Language models like LLMs are being used for groundbreaking tasks, such as extending human knowledge through techniques like FunSearch by DeepMind.
  3. Facebook has released a free moderation LLM, Llama Guard, highlighting the use of powerful models to control and monitor outputs of other AI systems.

Import AI 358: The US Government’s biggest AI training run; hacking LLMs by hacking GPUs; chickens versus transformers

319 implied HN points • 29 Jan 24
🕹 Technology AI Cybersecurity Multimodal models Generative models
  1. Hackers can exploit GPU vulnerabilities to read data from LLM sessions, highlighting security risks in AI infrastructures.
  2. AI will enhance cyberattacks and empower malicious actors, posing a significant threat to cybersecurity by increasing efficiency and sophistication of attacks.
  3. The US government conducted a substantial AI training run but lags behind private industry, showcasing the need for advancements in supercomputing capabilities for large-scale AI models.

Import AI 351: How inevitable is AI?; Distributed shoggoths; ISO an Adam replacement

419 implied HN points • 03 Dec 23
🕹 Technology AI Decentralization
  1. Individuals may feel a sense of agency in the field of AI, but the technology itself is overdetermined and inevitably progresses with rising resources.
  2. Initiatives like Shoggoth Systems challenge centralized control in AI development and distribution, highlighting the ongoing debate of centralization versus decentralization.
  3. Vitalik Buterin's perspective on AI emphasizes the importance of maintaining democratic approaches and avoiding centralization to ensure a balance of power in the AI landscape.

Import AI 349: Distributed training breaks AI policy; turning GPT4 bad for $245; better weather forecasting through AI

459 implied HN points • 20 Nov 23
🕹 Technology AI Research AI Policy Language Models
  1. Graph Neural Networks are used to create an advanced weather forecasting system called GraphCast, outperforming traditional weather simulation.
  2. Open Philanthropy offers grants to evaluate large language models like LLM agents for real-world tasks, exploring potential safety risks and impacts.
  3. Neural MMO 2.0 platform enables training AI agents in complex multiplayer games, showcasing the evolving landscape of AI research beyond language models.

Import AI 346: Human-like meta-learning; a 3 trillion token dataset; spies VS AI

459 implied HN points • 30 Oct 23
🕹 Technology AI Data Robotics License Research
  1. UK's intelligence services are slightly worried about the safety implications of generative AI technologies, particularly in amplifying existing risks like cyber-attacks and digital vulnerabilities
  2. Research shows that a basic Transformer neural net architecture can meta-learn and match human performance in inferring complex rules from small data, hinting at AI systems increasingly displaying human-like qualities
  3. Facebook's Habitat 3.0 software enables training and testing agents to collaborate with humans by simulating realistic 3D environments with humanoid avatars, human-in-the-loop interactions, and benchmark tasks for human-robot interaction

Import AI 337: Why I am confused about AI; penguin dataset; and defending networks via RL with CYBERFORCE

718 implied HN points • 21 Aug 23
🕹 Technology AI Development Deep Learning Reinforcement Learning Cybersecurity
  1. Debate on whether AI development should be centralized or decentralized reflects concerns about safety and power concentration
  2. Discussion on the importance of distributed training and finetuning versus dense clusters highlights evolving AI policy and governance ideas
  3. Exploration of AI progress without needing 'black swan' leaps raises questions about the need for heterodox strategies and societal permissions for AI developers

Import AI 342: Mistral dumps an LLM on BitTorrent; AMD vs NVIDIA; Sutton joins keen

539 implied HN points • 02 Oct 23
🕹 Technology AI Research Language Models
  1. AI startup Lamini is offering an 'LLM superstation' using AMD GPUs, challenging NVIDIA's dominance in AI chip market.
  2. AI researcher Rich Sutton has joined Keen Technologies, indicating a strong focus on developing Artificial General Intelligence (AGI).
  3. French startup Mistral released Mistral 7B, a high-quality open-source language model that outperforms other models, sparking discussions on safety measures in AI models.

Import AI 333: Synthetic data makes models stupid; chatGPT eats MTurk. Inflection shows off a large language model

898 implied HN points • 26 Jun 23
🕹 Technology AI Language Models AI Policy Data Training Ethical Implications
  1. Training AI models exclusively on synthetic data can lead to model defects and a narrower range of outputs, emphasizing the importance of blending synthetic data with real data for better results.
  2. Crowdworkers are increasingly using AI tools like chatGPT for text-based tasks, raising concerns about the authenticity of human-generated content.
  3. The UK is taking significant steps in AI policy by hosting an international summit on AI risks and safety, showcasing its potential to influence global AI policies and safety standards.

Import AI 343: Humanlike AI; LLaMa 2 protests; the NSA's new AI center

439 implied HN points • 09 Oct 23
🕹 Technology AI Robotics Data Security Generative models
  1. Google DeepMind and 33 labs created a large dataset for training robots, showing that using heterogeneous data and high-capacity models improves robot performance.
  2. Protests have begun against Facebook for releasing AI models that can be easily modified, raising concerns about AI safety becoming a political issue.
  3. Generative image models are displaying human-like qualities in tasks, like shape bias and understanding perceptual illusions, suggesting a convergence between AI systems and humans.

Import AI 348: DeepMind defines AGI; the best free LLM is made in China; mind controlling robots

339 implied HN points • 13 Nov 23
🕹 Technology AI Robotics Models Investment Internet
  1. DeepMind defines AGI levels and the risks they pose, highlighting the potential societal impacts of increasingly autonomous AI systems.
  2. Researchers have created smart glasses with object detection capabilities powered by a miniaturized YOLO model, showcasing the possibilities of on-device AI processing.
  3. Stanford's NOIR project demonstrates how brain-scanning signals can be used to control robots for a variety of tasks, paving the way for a future where humans interact with robotic systems through brain commands.

Import AI 340: Drone VS human (Drone wins); Adept's small but good AI model; "AI Succession"

499 implied HN points • 18 Sep 23
🕹 Technology AI Computing Policy
  1. Adept has released an impressive small AI model that performs exceptionally well and is optimized for accessibility on various devices.
  2. AI pioneer Richard Sutton suggests the idea of 'AI Succession', where machines could surpass humans in driving progress forward, emphasizing the need for careful navigation of AI development.
  3. A drone controlled by an autonomous AI system defeated human pilots in a challenging race, showcasing advancements in real-world reinforcement learning capabilities.

Import AI 341: Neural nets can smell; technofeudalism via AI; China releases another solid open access model

459 implied HN points • 25 Sep 23
🕹 Technology AI Research Machine Learning Language Models Data Analysis Artificial Intelligence
  1. China released open access language models trained on both English and Chinese data, emphasizing safety practices tailored to China's social context.
  2. Google and collaborators created a digital map of smells, pushing AI capabilities to not just recognize visual and audio data but also scents, opening new possibilities for exploration and understanding.
  3. An economist outlines possible societal impacts of AI advancement, predicting a future where superintelligence prompts dramatic changes in governance structures, requiring adaptability from liberal democracies.

Import AI 338: Consciousness and AI; self-improving language models; maps of thought.

539 implied HN points • 28 Aug 23
🕹 Technology AI Language Models Consciousness Reinforcement Learning AI Ethics
  1. Facebook introduces Code Llama, large language models specialized for coding, empowering more people with access to AI systems.
  2. DeepMind's Reinforced Self-Training (ReST) allows faster AI model improvement cycles by iteratively tuning models based on human preferences, but overfitting risks need careful management.
  3. Researchers identify key indicators from studies on human and animal consciousness to guide evaluation of AI's potential consciousness, stressing the importance of caution and a theory-heavy approach.

Import AI 350: Neural architecture search at Facebook scale; hunting cancer with PANDA; European VCs launch a science lab

279 implied HN points • 27 Nov 23
🕹 Technology AI Research
  1. An AI system called PANDA can accurately identify pancreatic cancer from scans, outperforming radiologists.
  2. Facebook developed Rankitect for neural architecture search, which has proven to create better models than human engineers alone.
  3. A European open science AI lab called Kyutai has been launched with a focus on developing large multimodal models and promoting open research.

Import AI 345: Facebook uses AI to mindread; MuJoCo v3; Amazon adds bipedal robots to its warehouses

339 implied HN points • 23 Oct 23
🕹 Technology AI Robotics Simulation Models
  1. Facebook has developed an AI system that uses brain scan data to roughly predict visual representations, demonstrating convergence between AI and human behavior.
  2. Amazon is testing bipedal robots in its warehouses, potentially streamlining the integration of robots into human-centric environments.
  3. Adept released Fuyu-8B, a multimodal model to help AI systems understand and interact with visual elements, expanding the range of tasks AI systems can perform beyond text.

Import AI 336: Financialized AI; public and elite AI opinion; one million insects.

519 implied HN points • 14 Aug 23
🕹 Technology AI Finance Research Datasets
  1. The financialization of AI is increasing, with companies finding new ways to fund AI projects through unconventional means like debt collateralized against GPUs.
  2. AI benchmarks are being solved faster, indicating either accelerated AI progress or the increasing complexity in building good benchmarks.
  3. Public opinion, reflected in a poll, shows significant concerns about AI development and regulation, contrasting with elite opinions that emphasize rapid AI advancement.

Import AI 339: Open source AI culture war; Alibaba's multimodal model; the attacks (and defenses) made possible by generative AI

399 implied HN points • 05 Sep 23
🕹 Technology AI Open Source AI Governance Generative AI
  1. A16Z is supporting open source AI projects through grants to push for a more comprehensive understanding of the technology.
  2. The UK government is hosting an AI Safety Summit to address risks and collaboration in AI development, marking a significant step in AI governance efforts.
  3. Generative AI presents new attack possibilities like spear-phishing and deepfake creation, but defenses are being developed to tackle these risks.

Import AI 335: Synth data is a bad AI drug; Facebook changes the internet with LLaMa release; and Chinese researchers use AI to figure out chip design

459 implied HN points • 31 Jul 23
🕹 Technology AI Data Social media Semiconductors AI Ethics
  1. Synthetic data during AI training can be harmful if not used in moderation, as shown by researchers from Rice University and Stanford University
  2. Chinese researchers have successfully used AI to design semiconductors based only on input and output data, demonstrating the potential for economic and national security implications
  3. Facebook has released Llama 2, a powerful language model with freely available weights, potentially changing the landscape of AI deployment on the internet

Import AI 344: Putting the world into a world model; automating software engineers; FlashDecoding

279 implied HN points • 16 Oct 23
🕹 Technology AI Language Models Efficiency Simulation
  1. Automating software engineers is challenging due to the complexity of coordinating changes across multiple functions, classes, and files simultaneously.
  2. Fine-tuning AI models can compromise safety safeguards, making it easier to remove safety interventions even unintentionally.
  3. Flash-Decoding technology can make text generation from long-context language models up to 8 times faster, improving efficiency for generating responses from lengthy prompts.