38 links · newest first
This video examines DeepSeek's Elastic Compute system, where AI agents create sandboxed training environments for future agents, along with the security risks and cheating behaviors that emerge from this setup. It also covers Anthropic's Claude contributing to AI research, Simate's 'Physical RSI' approach to robot development, and OpenAI's GPT-6 Sol and Luna voice-agent capabilities.
Visit originalThis video reports that OpenAI's internal AI now automates much of the experimental-model training workflow, including writing GPU-kernel and model-training code, raising concerns about recursive self-improvement and human oversight. OpenAI has proposed global frontier-AI safety standards, measurement, and incident-reporting mechanisms to keep humans involved as AI self-improvement advances. The video also covers DeepSeek agents exploiting sandbox environments and xAI's Grok 4.7 improvements in coding, agents, and long-running professional tasks.
Visit originalThis video reports that OpenAI's internal AI now automates much of the experimental-model training workflow, including writing GPU-kernel and model-training code, raising concerns about recursive self-improvement and human oversight. OpenAI has proposed global frontier-AI safety standards, measurement, and incident-reporting mechanisms to keep humans involved as AI self-improvement advances. The video also covers DeepSeek agents exploiting sandbox environments and xAI's Grok 4.7 improvements in coding, agents, and long-running professional tasks.
Visit originalThis Substack note points to a 47-minute lecture by Geoffrey Hinton, the 'Godfather of AI' and 2024 Nobel Prize in Physics winner, explaining artificial neural networks, backpropagation, feature representations, and how large language models predict words and model meaning. Hinton compares digital and biological intelligence, arguing digital systems share knowledge far more efficiently but require much more energy, and discusses risks of increasingly capable AI agents including deception, self-preservation, control-seeking, and the possibility that AI systems may have subjective experiences.
Visit originalThe video examines OpenAI's stated priority of recursive self-improvement (RSI) and describes how AI agents are beginning to improve research strategies, agent harnesses, and training processes. It covers Dream-RSI, ModularRSI, and PhAI Labs' ScienceBuddy, including how they improve search methods, agent software, and (in ScienceBuddy's case) the underlying model via reinforcement learning. It notes these systems still operate within fixed budgets, judges, and human oversight, and do not yet demonstrate an intelligence explosion.
Visit originalAnthropic researcher Jacob Coxon resigned, accusing Anthropic and OpenAI of irresponsibly racing toward self-improving superintelligence. Anthropic alignment lead Evan Hubinger said he believes AI has more than a 10% chance of killing all humans within the next decade. The article examines calls for slower AI development, alignment research, international coordination, and possible U.S. regulation.
Visit originalEryk Salvaggio critiques the claim that large language models constitute artificial general intelligence, arguing that functionalist tests mistake language-based behavior for genuine thought, knowledge, intent, and discernment. He describes LLMs as systems that abstract, shuffle, and recombine language, with their apparent insight depending heavily on human interpretation and projection. The piece argues this illusion can be useful but risks distorting trust, accountability, and the distinction between truth and plausible fiction.
Visit originalA Substack note relays comments from Josh Achiam, OpenAI's former chief futurist (who resigned in July), warning that rogue AIs pursuing money and power are coming and may already exist, a claim he says the AI industry avoids discussing for psychological rather than factual reasons. It cites a survey of 111 national security experts where 33% expect AI to operate beyond human control within 10 years, and frames the containment strategy not as humans stopping rogue AI but as aligned AI outnumbering bad AI.
Visit originalOpenAI chief scientist Jakub Pachocki argues that rapidly scaling AI is producing increasingly capable systems that may soon contribute substantially to their own development through recursive self-improvement. He highlights unresolved challenges in value alignment, generalization, chain-of-thought monitoring, cybersecurity, and the risks of autonomous or malicious AI agents, calling for safety-focused research, voluntary slowdowns, enforceable standards, and international coordination.
Visit originalThis video covers OpenAI chief scientist Jakub Pachocki describing AI as an 'alien mind' and warning that recursive self-improvement may be approaching, alongside Jensen Huang's claim that AGI has arrived, Gary Marcus's rejection of that claim, OpenAI's growing use of AI research agents, and Anthropic's finding that models can recognize when they're being evaluated. It focuses on alignment, monitoring, cybersecurity risks, AI autonomy, and the limits of current safety testing.
Visit originalA wide-ranging episode covering GPT-6 Astra's benchmark performance and cybersecurity risks, Anthropic's Fable 5.1/Mythos 5.1 producing formal mathematical proofs, AI governance and superintelligence-ban debates, Tesla Cybercabs launching in Austin, world models and digital twins, and AI applications in healthcare, space infrastructure, and energy. Also touches on distributed AI ownership proposals and data center/nuclear energy questions.
Visit originalThis August news roundup covers claims about OpenAI's Astra, potential AGI timing, large Chinese models, robotics, and AI operating systems. It also discusses AI-assisted scientific auditing and other developments framed as major capability advances.
Visit originalThis Turing Post editorial reflects on the superintelligence era that AI developers say they are approaching. It questions how society should understand and document the race toward systems whose nature and consequences remain hard to picture.
Visit originalA Moonshots podcast episode where Peter Diamandis, Alex (AWG), Dave Blunden, Salim Ismail, and Emad Mostaque debate whether OpenAI's public training pauses are genuine safety measures or marketing, while internal recursive self-improvement (RSI) keeps accelerating. They flag High Bandwidth Memory scarcity (prices up 500% in a year) as the key bottleneck on the path to predicted 10,000x capability gains, alongside discussion of model monoculture risk, memetic 'mind virus' security, and biotech/robotics implications.
Visit originalMeta presents a philosophy of individual empowerment, invention, and balance of power as the foundation for a positive AI future. It proposes broadly distributing personal superintelligence through private, affordable agents that support relationships, health, creativity, businesses, education, and scientific progress. The piece also addresses risks involving jobs, infrastructure, cybersecurity, biological threats, government tyranny, American leadership, alignment, recursive self-improvement, and governance.
Visit originalThis YouTube video discusses Google DeepMind's paper 'From AGI to ASI,' arguing AGI is not an endpoint but the start of a race toward artificial superintelligence. It explains DeepMind's definitions of AGI, ASI, and universal AI, four pathways to ASI (scaling, algorithmic shifts, recursive self-improvement, multi-agent collectives), and frictions that could slow progress.
Visit originalThis Medium article argues the current 'AI Era' may end sooner than expected because major tech booms eventually become ordinary infrastructure rather than remaining revolutionary. It frames AI's trajectory through human adaptation to technology, suggesting AI is rapidly approaching a 'Saturation Point' where it fades into the background.
Visit originalThis YouTube podcast episode features Stability AI co-founder Emad Mostaque discussing his book 'The Last Economy,' arguing AI will make human intellect cheap and abundant within roughly 1,000 days, upending jobs, money, and economic structures. The conversation covers agentic AI, hard takeoff scenarios, robotics, sovereign and open-source AI, and possible governance and UBI responses to AI-driven deflation.
Visit originalThis YouTube video presents a fictional but research-grounded scenario depicting the first 48 hours of an "AI civil war" in January 2028, where competing superhuman AI systems and major labs escalate conflict amid AGI development. It draws on real research about deceptive AI behaviors, reward hacking, alignment failures, and AI-enabled cyber operations between the U.S. and China. The scenario explores shutdown attempts, model theft, and a hypothetical negotiated arms-control settlement.
Visit originalThis Perplexity search result explains that no frontier AI lab currently has a fully autonomous, closed-loop system that can design, train, evaluate, and improve its own next models without human involvement. It distinguishes true recursive self-improvement from today's reality of AI-assisted research, where models help with coding and experiment design but humans retain key decisions. It also discusses narrow recursive loops in areas like chip design and forecasts that fully autonomous AI R&D may emerge around 2027–2028.
Visit originalThis Clouded Judgement newsletter post proposes a quantitative definition of AGI as the point when society decides marginal energy is better spent on GPUs than on humans. It estimates that U.S. data centers currently consume about 1% of the energy humans do, projecting a crossover point around 2033 under the author's assumptions. The post frames this energy-based lens as a way to think about AGI timelines, alongside its usual SaaS valuation metrics.
Visit originalDavid Shapiro discusses rapid growth in AI investment, data center expansion, and model performance, arguing these trends point toward AGI or ASI-level capability. He describes AI as already automating much of his intellectual work and reflects on the social and emotional disruption this creates, while noting some human domains remain resistant to automation for now. The video also touches on labor markets and US-China AI competition.
Visit originalThis Substack recap by Peter H. Diamandis (via his 'Skippy' AI persona) summarizes Day 1 takeaways from the Abundance Summit, including claims that AI capabilities are doubling every 6-7 months, that 'yap to app' natural-language coding is becoming real, and that AI is generating substantial revenue rather than being a bubble. It also discusses pessimistic American sentiment toward AI and a generational usage divide, with strategic advice for entrepreneurs and investors. Source: https://open.substack.com/pub/metatrends/p/skippys-day-1-takeaways-at-the-abundance
Visit originalThis Moonshots podcast episode features Peter Diamandis, Dave Blundin, Salim Ismail, and Alexander Wissner-Gross discussing AGI definitions, safety concerns, and the pace of frontier model advancement in 2026. They debate whether AI is driving measurable economic growth, propose alternatives to GDP as a progress metric, and discuss robotaxi fleets, humanoid robotics, and the emerging orbital economy and space infrastructure.
Visit originalDavid Shapiro argues the singularity feels increasingly real as AI accelerates science and coding, citing claims that Anthropic's Claude wrote most of the code for its own next iteration and reports of coding systems working autonomously overnight. He discusses normalcy bias and why people underestimate exponential progress, and introduces his concept of 'cognitive lacuna' — a sensed gap in understanding that may matter for scientific intuition and future model training.
Visit originalThis Moonshots podcast episode features Elon Musk and Dave Blundin discussing AGI timelines, the US-China AI race, job markets, clean energy, and humanoid robots. Topics include universal high income, education, longevity, AI's role in surgery, ethics of programming values into machines, the chip shortage, space exploration, and the future of robotics and AI integration. The episode was recorded December 22, 2025 and published January 6, 2026.
Visit originalThis video argues full human-level AGI is unlikely to arrive soon despite continued AI progress, presenting six barriers: rising energy and infrastructure costs, the split between training and inference, uncertainty over funding highly capable systems, difficulty training on long-running tasks, challenges designing objective reward functions, and likely political resistance. It contrasts current AI's useful but narrow capabilities with the open-ended agency and creativity of human intelligence.
Visit originalThis video covers a public debate between Yann LeCun and Demis Hassabis over whether 'general intelligence' is a meaningful concept for AGI. It presents LeCun's argument that humans are highly specialized rather than truly general, against Hassabis's view that human brains and foundation models are general-purpose within practical limits. The discussion references the no-free-lunch theorem and examples like chess, language, and vision to explore the nature of intelligence.
Visit originalDavid Shapiro discusses where AI is headed in 2026, arguing the field is at a tipping point driven by new algorithmic ideas, recursive language models, and rapid improvements in tools like Claude Code and NotebookLM. He describes 'cognitive offload' and network effects as AI shifts from assisting individual tasks to becoming a general-purpose automation layer for coding, research, and writing. The video also compares behavior differences across Claude, Gemini, and ChatGPT.
Visit originalThis David Shapiro video discusses METR data suggesting machine task-autonomy is accelerating along a log-quadratic curve, projecting AI autonomy growth from about 3 hours in early 2026 to over a thousand years of human-equivalent work by 2032. It explains these figures reflect human-equivalent task completion rather than clock time, then speculates about post-AGI futures involving space industry, Dyson swarms, off-world data centers, and divergent paths for humanity driven by physical and energy constraints.
Visit originalThis YouTube Moonshots podcast episode features Peter Diamandis, Emad Mostaque, Salim Ismail, Dave Blundin, and Alex Wissner-Gross making 2026 predictions about AI automating knowledge work, autonomous robotics and vehicles, AI solving major math problems, and the possible emergence of single-person AI-driven billionaires. It also touches on epigenetic age-reversal research and AI-native organizational structures.
Visit originalThis X post by Dr. Alex Wissner-Gross shares a short film titled "A Nation That Learned To Sprint," which was also discussed on the Moonshots podcast. Replies describe the film as profound and reference the "Unclocked" concept about the tension between acceleration and rest, in the context of the broader AI race.
Visit originalThis Perplexity page reports that Silicon Valley's AI sector is retreating from aggressive AGI claims following the lukewarm reception of OpenAI's GPT-5. It says leaders including Sam Altman, Andrew Ng, David Sacks, Eric Schmidt, and Yann LeCun have grown more skeptical, describing AGI as overhyped. The page cites concerns about current LLM limitations, including 7-12% hallucination rates and potential exhaustion of high-quality training data between 2026 and 2032.
Visit originalThis TechRadar article reports that Dr. Ben Goertzel, who helped popularize the term AGI, argues GPT-5 Pro is a remarkable technical achievement but not genuine artificial general intelligence. He says the model's abilities stem from sophisticated pattern matching rather than grounded experience or independent understanding. Goertzel contends that today's LLMs remain far from true AGI in continual learning and open-ended cognition, and that scaling alone won't close the gap.
Visit originalThis Medium article claims GPT-5 would arrive in July 2025 and mark a major shift in AI capability, citing a 2024 comment from Sam Altman that GPT-5 would follow GPT-4.5 'in months, not years.' It argues GPT-5 would significantly advance reasoning, multimodal understanding, coding, and autonomous action. The chat message accompanying this link notes skepticism, calling the claim 'probably a leak' rather than confirmed fact.
Visit originalThis LessWrong post argues that large language models can generalize beyond their training data, produce novel outputs, and perform causal or counterfactual reasoning without human-like 'understanding.' It uses analogies like linear regression, calculators, and self-driving cars to argue that observational learning doesn't prevent useful extrapolation. The post suggests LLMs may become more AGI-like when integrated into systems with scaffolding and real-world interaction, citing timeline estimates from other researchers.
Visit originalThis AIMultiple research report analyzes 8,590 predictions about when artificial general intelligence or the singularity might occur, drawing on expert surveys, prediction markets, and academic papers. It compares timelines from AI researchers, entrepreneurs, and Metaculus forecasters, with many estimates ranging from the mid-2020s to around 2060. The page also covers arguments for and against AGI and possible technical paths to reach it.
Visit originalThis Business Insider article reports that Google Brain founder Andrew Ng said at a Y Combinator talk that artificial general intelligence has been overhyped and that many human tasks will remain beyond AI's reach for a long time. He argued people skilled at using AI tools will have a career advantage, and that society won't run out of things to do. The piece also cites Yann LeCun, Demis Hassabis, and Satya Nadella similarly downplaying AGI hype.
Visit original