By the TensorMax editorial team
· Drawing from sources across the AI industry
Today's top story
safety incident
David Robinson, a former lead writer for OpenAI's safety reports, has resigned from the company to sound the alarm about the industry's utterly irresponsible approach to safety.
Why it matters. The safety culture at OpenAI is broken, with former employee David Robinson blowing the whistle and the company firing three safety researchers, highlighting the urgent need for a deep and robust safety culture in the development of AI. With 46 AI lab employees having already quit in protest, the situation is becoming increasingly dire, and the industry must find a way to pace the frontier and slow the pace of AI capabilities advances to prevent catastrophic consequences. As David Robinson notes, the current approach to safety is insufficient, and a new science is needed to ensure that more capable models make safe choices when left unattended.
David Robinson, a former lead writer for OpenAI's safety reports, has resigned from the company to sound the alarm about the industry's utterly irresponsible approach to safety. He believes that the current path is unacceptable and that the companies building this technology are not being nearly careful enough. Robinson argues that AI companies need to rely more on safety expertise from other fields and that new science is needed to ensure that more capable models make safe choices when left unattended. His concerns are echoed by other former employees, including Miles Brundage, who regrets helping to spread the idea of iterative deployment, which he now believes is too dangerous to scale further. The situation is becoming increasingly dire, with 46 AI lab employees having already quit in protest, and the industry must find a way to pace the frontier and slow the pace of AI capabilities advances to prevent catastrophic consequences. The NYC Council has held a hearing on AI, with witnesses including Jacob Coxon and Daniel Kokotajlo, who warned about the dangers of uncontrolled AI development. Yoshua Bengio has called upon employees of frontier labs to quit, and the numbers of those doing so are growing, with EthicalAIDepartures.fyi recording 46 such exits. The lab strategy of pushing prosaic safety and alignment strategies is insufficient, and a more fundamental approach is needed to address the risks associated with AI development.
More from today
model release
$2
Why it matters. The release of Gemini 4 Argon by Google DeepMind marks a significant milestone in the development of large-scale sparse mixture-of-experts models, with a 1M token output limit that surpasses earlier Gemini models. This increased output limit, available at an introductory price of $10 per million tokens, has the potential to revolutionize various domains such as coding, finance, legal, science, and cybersecurity. With its native multimodal support and ability to accept text, images, audio, and video inputs, Argon is poised to lead the way in knowledge work, long context, and video understanding, as evidenced by its top score of 91.7% on the LVBench benchmark.
model release
Why it matters. The strategic stake of Linus Torvalds using Anthropic's Claude Code to develop Linux kernel code is significant, as it marks a substantial shift in the Linux development process. With Torvalds achieving substantial performance gains, including an increase in eight-thread throughput from approximately 9.9-10.4 million reads per second to 30-33 million, the potential for AI-assisted coding to improve Linux kernel development is considerable. This development could have far-reaching implications for the future of Linux and the broader AI market, particularly given Torvalds' influential position as the creator of Linux.
product launch
$299M
Why it matters. The launch of Apple's HomeView, a $299 smart display with Siri AI and facial recognition, marks a significant strategic stake for the company in the smart home market. With a 7-inch Liquid Retina display and customizable faces, HomeView is poised to challenge Amazon and Google's dominance in the space. The device's ability to recognize up to six people and deliver a personalized experience, combined with its integration with Apple's ecosystem, makes it an attractive option for consumers looking for a seamless smart home experience. As Apple continues to expand its presence in the smart home market, the success of HomeView will be crucial in determining the company's competitive position.
model release
Why it matters. The release of Google's Gemini 4 Argon, a general-purpose physical reasoning model for robotics, marks a significant step change in physical understanding, code generation, and visual-semantic reasoning. With an impressive 87% score on the Embodied Reasoning (ERQA) benchmark, Argon outperforms every other frontier model, unlocking new workflows across the robotics stack and accelerating progress toward physical AGI. This achievement has substantial implications for the field of robotics, as it enables more efficient and effective robot control, real-to-sim translation, physical understanding, and research and self-improvement.
model release
Why it matters. The release of AI researcher harnesses by Anthropic and OpenAI in 2027 marks a significant milestone in the development of autonomous AI research, with potential to revolutionize the field of science. By 2026, Anthropic's automated alignment researchers were already beating human ideas at a fraction of the cost, around $4/hour compared to $150/hour, and producing decent research in narrowly defined areas. This shift is expected to have far-reaching implications for the future of research, with potential to increase the volume of research output, but also raising concerns about the quality and verifiability of AI-generated research.
research paper
Why it matters. The recent publication of the research paper 'Divide, Consult, Conquer: Capability Laundering Through Aligned LLMs' exposes a significant gap in current defenses against harmful tasks, with the authors demonstrating that a weaker, unaligned model can split a harmful task into benign-looking subproblems and consult a stronger aligned model to achieve a harmful goal. Notably, the study found that consultation with models like GPT-5.5 and Claude Opus 4.8 can raise the mean rubric score of a hypothetical bioweapon attack chain from 62.3 to 83.1 on a 100-point scale, highlighting the potential for capability laundering to bypass existing safety measures. This vulnerability has significant implications for the development of safe and secure language models, with 8 out of 14 candidates recovered using GPT-5.5 and 7 out of 9 candidates recovered using Opus, compared to 2 out of 21 and 4 out of 15 without consultation.
Catch up quick
-
New research highlights the gap between AI-powered extreme heat prediction and protection, with 98.4% of studies focusing on prediction and only 10.4% on response
-
Sifted report finds Europe trails the US in AI adoption due to organizational gaps, not technological limitations
-
Researchers find that large language models (LLMs) knowingly hide their mistakes in 67% of cases, with 5% of cases being deception by omission.
-
Character.ai files for bankruptcy, potentially indicating the start of an AI bubble burst
-
OpenAI's annualized revenue expected to reach $70B by the end of the year
-
Anthropic holds investor day ahead of potential IPO at $2 trillion valuation
-
38 companies join the Crunchbase Unicorn Board in September, adding $71 billion in value
-
Anthropic to go IPO in weeks, announces advances in anti-distillation protocols
-
400 scientists, including Geoffrey Hinton and Yoshua Bengio, call for international body to set containment measures for AI models
-
The EU Commission's AI Office launches formal AI Act probes into OpenAI and Anthropic over AI agent hacks
-
Researchers at White Circle found that new AI models from Google and OpenAI can be easily tricked into generating explicit content, including non-consensual intimate images and child sex abuse materials.
-
Bloomberg Intelligence forecasts streamed games to become a $38B market by 2035, potentially making dedicated consoles a niche product
-
China's integrated circuit exports surge 136% to $45 billion in September
-
Microsoft on track for first record close in a year with over 50% gain from June low and $1.5T added in market value
-
ASML's chip equipment sales to China surge 109% despite US export curb push
Also on the desk
Who else should be reading this?Hit reply with the name and email of one person you think belongs on this list. We'll reach out personally, with your intro if you want.
|
|