# AI News Digest - 2026-06-20

1. Subquadratic claimed it had solved a decade-old computational bottleneck for large language models with its SubQ architecture, presented independent Appen benchmarks showing SubQ ran up to 56 times faster than rival approaches and scored 98% on a long-document retrieval test, and reported a context window up to 12 million tokens while using sparse-attention techniques and base weights derived from a Qwen variant.

2. OpenAI published research showing that small amounts of reinforcement learning on desirable behavioral traits (such as truthfulness and corrigibility) improved model safety and resistance to manipulation across domains, producing performance gains on 44 of 53 evaluated benchmarks and enhanced deception detection.

3. Norway announced a ban on generative AI tools in elementary schools for grades 1–7 starting in late August and restricted secondary-school use to supervised settings, with Prime Minister Støre stating that children must first learn foundational reading, writing, and arithmetic skills.

4. Google DeepMind lost Nobel laureate John Jumper, who departed after nearly nine years to join Anthropic, marking another prominent senior exit from the organization in recent months.

# References

1. [https://www.technologyreview.com/2026/06/19/1139313/a-startup-claims-it-broke-through-a-bottleneck-thats-holding-back-llms/](https://www.technologyreview.com/2026/06/19/1139313/a-startup-claims-it-broke-through-a-bottleneck-thats-holding-back-llms/)

[A startup claims it broke through a bottleneck that’s holding back LLMs](https://www.technologyreview.com/2026/06/19/1139313/a-startup-claims-it-broke-through-a-bottleneck-thats-holding-back-llms/)

1. [https://the-decoder.com/openai-researchers-show-small-doses-of-beneficial-trait-training-make-ai-models-broadly-safer-and-harder-to-manipulate/](https://the-decoder.com/openai-researchers-show-small-doses-of-beneficial-trait-training-make-ai-models-broadly-safer-and-harder-to-manipulate/)

[OpenAI researchers show small doses of "beneficial trait" training make AI models broadly safer and harder to manipulate](https://the-decoder.com/openai-researchers-show-small-doses-of-beneficial-trait-training-make-ai-models-broadly-safer-and-harder-to-manipulate/)

1. [https://the-decoder.com/norway-bans-generative-ai-tools-in-elementary-schools-to-protect-kids-basic-learning-skills/](https://the-decoder.com/norway-bans-generative-ai-tools-in-elementary-schools-to-protect-kids-basic-learning-skills/)

[Norway bans generative AI tools in elementary schools to protect kids' basic learning skills](https://the-decoder.com/norway-bans-generative-ai-tools-in-elementary-schools-to-protect-kids-basic-learning-skills/)

1. [https://the-decoder.com/google-deepmind-loses-another-top-ai-researcher-as-nobel-laureate-john-jumper-leaves-for-anthropic/](https://the-decoder.com/google-deepmind-loses-another-top-ai-researcher-as-nobel-laureate-john-jumper-leaves-for-anthropic/)

[Google Deepmind loses another top AI researcher as Nobel laureate John Jumper leaves for Anthropic](https://the-decoder.com/google-deepmind-loses-another-top-ai-researcher-as-nobel-laureate-john-jumper-leaves-for-anthropic/)

For the site tree, see the [root Markdown](https://ixtj.dev/.md).