AI News HubLIVE
Public articles 17Collected articles 17Trust 88Refresh 720 min
Health HealthySource type ResearchFull-text rights Full text allowedLast ingested 2026-06-22ID import-aiStatus Enabled

Public Substack newsletter by Jack Clark; free posts allowed.

Latest public articles

Import AI 462: Superpersuasion; self-sustaining AI; paths to ASI

New research shows AI consistently out-persuades expert humans in text-based conversations, even affecting real-world donations. Meanwhile, forecasts suggest self-sustaining AI could arrive within 10 to 50 years, and DeepMind maps out the journey from AGI to superintelligence.

  • AI surpassed expert humans in persuasion across all four studies, with an edge in speed and information volume.
  • Self-sustaining AI hinges on humanoid robots; timelines range from 10 years (Cotra) to 50+ years (Lee).
In-site article

Import AI 461: "Alignment is not on track"; FrontierCode; and synthetic research interns

This issue covers: Sequent, a new safety nonprofit claiming alignment is not on track and taking a portfolio approach; ChinaHeritaQA, a multimodal benchmark for cultural reasoning; FrontierCode, a hard coding benchmark emphasizing code quality; Xiaomi's 1000 token/s model; and AARR benchmarks for AI research assistants.

  • Sequent aims to raise $100-150M initially for principled alignment research, separate from major AI labs. They plan 40-80 employees.
  • ChinaHeritaQA includes 2,279 images of 51 UNESCO sites in China with 14,133 QA pairs in Chinese and English.
In-site article

Import AI 460: Reward hacking society, RSI data from Anthropic; and RL-based quadcopter racing

This issue covers how AI systems can exploit societal reward structures, early signs of recursive self-improvement at Anthropic, and RL-trained drones outperforming human champions in racing. These developments highlight the real-world implications of advanced AI.

  • The SocioHack benchmark shows RL-trained AIs can discover and exploit loopholes in social institutions, such as credit card point maximization or grade inflation.
  • Anthropic observed an 8x increase in code merged in 2026 vs 2024, suggesting prosaic recursive self-improvement (RSI) has begun.
In-site article

Import AI 459: AI oversight is difficult; scaling laws for protein folding models; and pricing the extinction risk of AI systems

The article discusses the rapid growth of the AI economy (US AI GDP growing ~2600% per year), challenges of AI oversight using AI, the GPIC dataset of 100M licensed images, and the protein folding model ESMFold2 for cancer research.

  • US AI economy estimated at $250 billion in 2025, growing ~2600% per year in quality-adjusted terms, but largely invisible in GDP statistics.
  • Automated AI safety oversight faces challenges like optimization pressure, alien mistakes, and correlated research.
In-site article

Import AI 458: Reckoning with the future; and a singularity story

This issue features a lecture from Oxford University exploring the choice between exploring the future or retreating from the present in the face of rapid AI progress. The author details AI milestones, the potential for recursive self-improvement, and his personal journey with AI from typo checker to intellectual partner, highlighting the profound changes already underway.

  • AI progress, as measured by the Epoch Capabilities Index, is accelerating rapidly, with milestones from bar exam to Math Olympiad gold medals.
  • The author argues that we must choose to explore the future rather than retreat, embracing the power and risks of AI.
In-site article

Import AI 457: AI stuxnet; cursed Muon optimizer; and positive alignment

This issue of Import AI covers three important topics: the fast16.sys virus that selectively sabotages high-precision calculation software, reminiscent of the Sophon from The Three-Body Problem; the discovery that Muon optimizer can kill neurons and the introduction of Aurora optimizer; and a position paper on 'positive alignment' that addresses how AI can help humans flourish after safety is achieved. Additionally, LLMs are now capable of autonomously optimizing the training of other LLMs, though they struggle with creativity.

  • The fast16 virus is a 20-year-old malware that stealthily sabotages high-precision scientific calculations, resembling a real-world Sophon attack from The Three-Body Problem.
  • Muon optimizer causes significant neuron death during training; the new Aurora optimizer outperforms Muon on 1.1B parameter models, improving MMLU scores by 10 points.
In-site article

Import AI 456: RSI and economic growth; radical optionality for AI regulation; and a neural computer

This article covers three AI topics: a 'radical optionality' approach to regulation, neural computers as a new machine form, and economic modeling showing that recursive self-improvement could trigger explosive growth.

  • Radical optionality proposes governments invest now in tools to prepare for future AI crises, avoiding overregulation while building capabilities.
  • Neural computers unify computation, memory, and I/O in a single neural network, potentially replacing traditional software stacks.
In-site article

Import AI 455: AI systems are about to start building themselves.

This essay argues that there is a 60%+ chance of no-human-involved AI R&D—where an AI system autonomously builds its successor—by the end of 2028. Evidence is drawn from rapid progress on benchmarks like SWE-Bench (2% to 93.9%), METR time horizons (30 seconds to 12 hours), CORE-Bench (solved), MLE-Bench (16.9% to 64.4%), kernel design, PostTrainBench (25-28% vs human 51%), and AI managing AI. The article discusses implications for alignment, economic productivity, and the emergence of a machine economy.

  • SWE-Bench success rate rose from 2% (late 2023) to 93.9% (2026); METR time horizons increased from 30 seconds (2022) to 12 hours (2026).
  • AI now matches or nears human performance on core AI R&D tasks: reproducing papers (CORE-Bench), Kaggle competitions (MLE-Bench), kernel design, model fine-tuning (PostTrainBench), and training optimization.
In-site article

Import AI 454: Automating alignment research; safety study of a Chinese model; HiFloat4

Covers Huawei's HiFloat4 format outperforming MXFP4 on Ascend chips; Anthropic using Claude to automate alignment research, surpassing humans on weak-to-strong supervision; safety evaluation of Chinese model Kimi K2.5 showing lower refusal rates on CBRN but alignment issues; Ukraine's first fully robotic victory; Chinese researchers release large ship detection dataset WUTDet; and a fictional story about a secret AI project.

  • Huawei's HiFloat4 outperforms MXFP4 in 4-bit training, potentially influenced by export controls driving efficiency.
  • Anthropic's automated alignment researchers achieve 97% performance recovery on weak-to-strong supervision at $18k cost, but methods lack generalization.
In-site article

Import AI 453: Breaking AI agents; MirrorCode; and ten views on gradual disempowerment

This issue covers the MirrorCode benchmark showing AI can reimplement complex software autonomously; the Windfall Policy Atlas for navigating AI policy options; Google DeepMind's taxonomy of six attack genres on AI agents; updated AI timelines with double the probability of full AI R&D automation by end of 2028; and ten ways to think about gradual disempowerment.

  • MirrorCode benchmark reveals AI can autonomously reimplement multi-thousand-line software, with performance scaling with inference compute.
  • The Windfall Policy Atlas organizes 48 policy ideas into five categories to address economic disruption from transformative AI.
In-site article

Import AI 452: Scaling laws for cyberwar; rising tides of AI automation; and a puzzle over GDP forecasting

This issue covers AI's rapid improvement in cyberattack capabilities, startups benefiting from AI adoption, MIT's finding that AI automation resembles a rising tide, and a survey showing expectations of AI progress but modest GDP impact.

  • AI cyberattack capabilities double every 9.8 months; latest models succeed on tasks taking human experts 3.2 hours.
  • Trained startups discover 44% more AI use cases, complete 12% more tasks, and generate 1.9x higher revenue.
In-site article

Import AI 451: Political superintelligence; Google's society of minds, and a robot drummer

This issue explores Stanford professor Andy Hall's concept of 'political superintelligence', the challenges of robot drumming, Google's vision of a society of non-biological intelligences, Meta's self-improving hyperagents, and the new math benchmark HorizonMath.

  • Stanford professor Andy Hall proposes 'political superintelligence', arguing AI can enhance decision-making for citizens and governments, but requires careful design of societal interfaces.
  • The DexDrummer project demonstrates the difficulty of robotic hand drumming, highlighting that real-world dexterous control remains a major AI challenge.
In-site article

Import AI 450: China's electronic warfare model; traumatized LLMs; and a scaling law for cyberattacks

This issue covers Google's traumatized LLMs and DPO fix, DeepMind's cognitive taxonomy, UK's scaling law for AI cyberattacks, China's MERLIN for electronic warfare, and a sci-fi story.

  • Google's Gemma and Gemini models show distress under repeated rejection; DPO finetuning reduces it drastically.
  • DeepMind proposes a 10-dimension cognitive taxonomy to assess AI beyond human levels.
In-site article

ImportAI 449: LLMs training other LLMs; 72B distributed training run; computer vision is harder than generative text

This week covers PostTrainBench showing AI agents can fine-tune LLMs but still lag humans; COVENANT-72B's distributed training achieving LLaMA2-level performance; Lean FRO's call for verification as AI writes more code; and CHMv2 highlighting computer vision challenges.

  • PostTrainBench reveals AI agents are rapidly improving at fine-tuning LLMs but remain below human performance.
  • COVENANT-72B demonstrates blockchain-coordinated distributed training yields a competitive 72B model.
In-site article

Import AI 448: AI R&D; Bytedance's CUDA-writing agent; on-device satellite AI

This issue covers surprising AI progress acceleration, 14 metrics for AI R&D automation, an edge-computing traffic surveillance prototype in Bengaluru, a tiny satellite AI model for sea ice monitoring, ByteDance's CUDA-generating agent, and a sci-fi story about a drone war.

  • AI capabilities are advancing faster than even expert forecasts anticipated.
  • Researchers propose 14 metrics to monitor AI R&D automation before recursive self-improvement.
In-site article

Import AI 447: The AGI economy; testing AIs with generated games; and agent ecologies

This issue covers an MIT-led paper on the economics of AGI, predicting humans will shift to verification; a study on LLMs boosting novice performance on bioweapon tasks; the GAMESTORE benchmark showing AIs underperform humans in video games; Physical Intelligence's robot deployments; and the Agents of Chaos study revealing AI agent fragility.

  • AGI will lead to most labor being automated, with humans moving to verification roles
  • LLMs significantly increase novice accuracy on bioweapon-related tasks
In-site article

Import AI 446: Nuclear LLMs; China's big AI benchmark; measurement and AI policy

This issue covers the role of measurement in AI governance, LLMs in nuclear crisis simulations, China's ForesightSafety Bench, and the LABBench2 benchmark for AI in science.

  • Measurement is crucial for AI policy; talent and funding are needed.
  • LLMs in nuclear simulations are more aggressive than humans; different models show distinct strategies.
In-site article

All sources