• Surging AI adoption in education:
  • Broad international adoption has occurred among students, with over 80%–94% of university and school students reporting AI tool usage across various countries.
  • A major study tracked 26,811 secondary school students (aged 12–18) in China between January 2023 and June 2025 to evaluate the impact of LLMs (such as Doubao and DeepSeek) on academic performance.
• The "AI learning penalty" and performance divergence:
  • Homework improvements: AI users saw average homework scores rise by 18%, while completion time fell from 64 minutes down to 45 minutes.
  • Exam declines: When tested independently in exams without AI assistance, these same students scored 20% lower than peers who did not use AI.
  • Decoupling of metrics: High homework grades historically predicted exam success; with AI, top homework scores now correlate with worse exam performance.
• Study behavior and usage methods dictate outcomes:
  • The exam penalty was heavily concentrated among students who used AI to rush assignments and copy-paste answers.
  • Students who spent equivalent study time while utilizing AI—using models as personal tutors for conceptual explanations rather than answer shortcuts—retained solid exam results.
  • A complementary Middlebury College study on undergraduates found that when used actively to learn unfamiliar material, AI tools improved both immediate and long-term test performance.

Hacker News Discussion

• Crowding out effort vs. force amplification:
  • Commenters discussed whether AI amplifies capabilities or simply reduces the cognitive struggle essential for learning.
  • Citations from the paper highlighted that over time, students learn how to take shortcuts, which eventually eliminates high-effort study sessions (spending over 65 minutes on homework vanished after months of adoption).
• Demographics, agency, and meta-learning:
  • Users debated if AI benefits advanced, self-driven learners (e.g., graduate students) while harming middle/high schoolers who lack academic agency and study primarily out of obligation.
  • Some countered that the study showed high-achieving students also suffered substantial negative learning effects when turning to AI shortcuts.
• Critiques of traditional academic assessments:
  • Commentators argued that standard exams often measure compliance, test preparation time, or rote memorization rather than deep comprehension or aptitude.
  • Educators noted that AI exposes structural weaknesses in university and secondary assessment methods, which have failed to evolve pedagogical standards alongside technology.

大多数人认为AI应该严格限制在特定任务和数据集内，以避免信息污染和边界模糊，但作者认为AI应该能够跨渠道学习并整合不同来源的信息。这挑战了人们对AI应用范围和数据隔离的传统认知，暗示未来AI将更像是具有广泛知识背景的团队成员。

大多数人认为AI模型主要是通过训练数据学习，而非通过实际业务操作进行学习。但作者暗示Claude正在通过企业部署过程中实时学习业务流程和决策逻辑，这种学习方式挑战了传统AI模型的训练范式，暗示AI可能正在从静态训练向动态学习转变。

大多数人认为AI模型主要是在受控环境中学习和训练，但这里暗示Claude正在通过实际业务操作直接学习企业运作模式，这种在真实商业环境中持续学习的方式挑战了传统AI训练方法的封闭性和局限性，暗示AI可能正在向自主学习和适应的方向发展。

Whole paragraph is good on the difference between "data processing" by AI and human intelligence/understanding/wisdom. Really intrigued here by the idea that forgiveness and fidelity are keys to learning.

We may want to focus on this approach - potentially more appealing to faculty

should be moved to forefront, before technical proficiency

Very interested in how this is done. Metacognition needs to be foregrounded in learning with AI, so learners can determine if the tech is helping and how it is doing so.

大多数人认为AI可以替代人类学习过程，但作者认为跳过调试和错误经验会阻碍隐性知识的形成，导致关键能力无法传承。这与AI可以完全替代人类学习的普遍认知相悖。

Moving things around and arranging thoughts and ideas in an essay is an essential part of the writing process.

大多数人认为AI的创造力主要来自模型规模和计算能力的提升，而作者认为缺乏持续学习和多样性机制将限制AI的真正创造力。这一观点挑战了主流AI发展路径，暗示技术规模扩张本身不足以实现真正的科学创新。

文章以'文件柜'的比喻生动地说明了当前AI系统的局限性。即使上下文窗口不断扩大，本质上仍然只是更大的文件柜。真正的突破是让模型在部署后继续执行训练时的核心能力：压缩、抽象和学习。这个观点挑战了当前AI发展的主流方向，提出了一个令人深思的问题：我们是否在追求错误的解决方案？

通常认为 AI 工具缺乏学习和适应能力，但作者提出 AI 代理可以通过团队的使用和反馈不断改进，这与主流观点中对 AI 学习能力的看法相悖。

大多数人可能认为 AI 工具的改进主要依赖于开发者，但作者强调 agents 的记忆和对话指导能力，使得它们在使用过程中不断改进。

令人惊讶的是：研究表明学习生成图像实际上能显著提升细粒度视觉理解能力，这一发现挑战了传统认知，即理解能力与生成能力应该是分离的，这为AI模型设计提供了全新的思路。

• [ ] return

https://en.wikipedia.org/wiki/Yann_LeCun

WOW! I've been out of the classroom for quite awhile and never considered this scenario regarding AI. This hit a nerve in me as I'm sure it will with many. I get it!

How do we respond and mitigate the isolation, the loss of human dialogue, mentorship and connection?

AI Tools for each phase of analysis

Pulling back to this post - https://wiobyrne.com/embracing-the-violin-of-cognitive-amplification/

Anshumali's prime research work on SLIDE algorithms.

MAPPING SOCIAL CHOICE THEORY TO RLHF Jessica Dai and Eve Fleisig ICLR Workshop on Reliable and Responsible Foundation Models 2024

Nice overview of how social choice theory has been connected to RLHF and AI alignment ideas.

• Page 17: Top 5 most important factors for creating an effective teaching and learning ecosystem: Having a strong leadership and vision (45%) is the #1 (next highest is 15%)
• Page 20: *83% of higher education respondents said that it was important for institutions to provide studens with skills-based learning alongside their academic education. *
• Page 26: Participants identified several challenges in fostering a a culture of lifelong learning for professionals, including: 89% Clear learning objectives
• Page 7: Real-world experiential and work-based learning are no longer fringe; 4 in 5 see these as essential.

( ~ 6:25-end )

Steps for designing a reading plan/list: 1. Pick a topic/goal (or question you want to answer) & how long you want to take to achieve this. 2. Do research into the books necessary to achieve this goal. Meta-learning, scope out the subject. The number of books is relative to the goal and length of the goal. 3. Find the books using different tools such as Google & GoodReads & YouTube Recommendations (ChatGPT & Gemini are also useful). 4. Refine the book list (go through reviews, etc., in Adlerian steps, do an Inspectional Read of everything... Find out if it's truly useful). Also order them into a useful sequence for the syntopical reading project. Highlight the topics covered, how difficult they are, relevancy, etc. 5. Order the books (or download them)

Reminds me a bit of Scott Young's Metalearning step, and doing a skill decomposition in van Merriënboer et al.'s 10 Steps to Complex Learning

Reply to Nidhi Sachdeva: Nidhi Sachdeva, PhD Just went through the micro-lesson itself. In the context of teachers using to generate instruction examples, I do not argue against that. The teacher does not have to learn the content, or so I hope.

However, I would argue that the learners themselves should try to come up with examples or analogies, etc. But this depends on the learner's learning skills, which should be taught in schools in the first place.

Reply to Nidhi Sachdeva: I agree mostly, but I would advice against using AI for this. If your brain is not doing the work (the AI is coming up with the story/analogy) it is much less effective. Dr. Sönke Ahrens already said: "He who does the effort, does the learning."

I would bet that Cognitive Load Theory also would show that there is much less optimized intrinsic cognitive load (load stemming from the building or automation of cognitive schemas) when another person, or the AI, is thinking of the analogies.

https://www.linkedin.com/feed/update/urn:li:activity:7199396764536221698/

Programming with LLM = less stress

In a study, students who learned to code with AI made more progress during training sessions, had significantly higher correctness scores, and retained more of what they learned compared to students who didn't learn with AI.

T. Herlau, "Moral Reinforcement Learning Using Actual Causation," 2022 2nd International Conference on Computer, Control and Robotics (ICCCR), Shanghai, China, 2022, pp. 179-185, doi: 10.1109/ICCCR54399.2022.9790262. keywords: {Digital control;Ethics;Costs;Philosophical considerations;Toy manufacturing industry;Reinforcement learning;Forestry;Causality;Reinforcement learning;Actual Causation;Ethical reinforcement learning}

Sometimes, search is not rooted in discovery of a correct answer to a question. It's about exploration. Serendipity through search. Think Michael Lewis, Malcolm Gladwell, and Latif Nasser from Radiolab. The randomizer on wikipedia. A risk factor of where things trend with advanced AI in search is an abandonment of meaning making through exploration in favor of a knowledge-level pursuit that lacks comparable depth to more exploratory experiences.

• for: canonical unit, collaborative commons - missing part - open learning commons, question - process trap - natural capital

• comment

  • in this context, indyweb and Indranet are not the canonical unit, but then, it seems the model is fundamentally missing the functionality provided but the Indyweb and Indranet, which is and open learning system.
  • without such an open learning system that captures the essence of his humans learn, the activity of problem-solving cannot be properly contextualised, along with all of limitations leading to progress traps.
  • The entire approach of posing a problem, then solving it is inherently limited due to the fractal intertwingularity of reality.

• question: progress trap - natural capital

  • It is important to be aware that there is a real potential for a progress trap to emerge here, as any metric is liable to be abused

Wang et. al. "Scientific discovery in the age of artificial intelligence", Nature, 2023.

A paper about the current state of using AI/ML for scientific discovery, connected with the AI4Science workshops at major conferences.

(NOTE: since Springer/Nature don't allow public pdfs to be linked without a paywall, we can't use hypothesis directly on the pdf of the paper, this link is to the website version of it which is what we'll use to guide discussion during the reading group.)

Good points & questions to raise

• Alphago
  • first version took months of Google UK software developers to program. It won the world Go championship.
  • Alphago Master played itself without ever watching a human player. It beat the first Alphago version after 3 days of playing itself.
  • In 21 days, it beat Alphago version one a thousand to zero.

LeBlanc, D. G., & Lee, G. (2021). General Deep Reinforcement Learning in NES Games. Canadian AI 2021. Canadian Artificial Intelligence Association (CAIAC). https://doi.org/10.21428/594757db.8472938b

[29] AI - Deep Learning

No new physics and no new mathematics was discovered by the AI. The AI did however deduce something from the existing math and physics, that no one else had yet seen. Skynet is not coming for us yet.

Here’s a lovely blog by Jason Wei that pulls together 137 examples of ’emergent abilities of large language models’. Emergence is a phenomenon seen in contemporary AI research, where a model will be really bad at a task at smaller scales, then go through some discontinuous change which leads to significantly improved performance.

i don't really understand this: in z-code thre are tasks that other competitive softwares would need to restart all over again while z-code can do it without restarting...

ReconfigBehSci on Twitter: ‘Now #scibeh2020: Pat Healey from QMU, Univ. Of London speaking about (online) interaction and miscommunication in our session on “Managing Online Research Discourse” https://t.co/Gsr66BRGcJ’ / Twitter. (n.d.). Retrieved 6 March 2021, from https://twitter.com/SciBeh/status/1326155809437446144

And the problem is that, even human beings are not very sensitive to how this can be done well. Because there is such diversity in human cultures, preferences, and norms, deciding whose values to prioritise is problematic.

Same is true of reinforcement learning algorithms.

Yang, Scott Cheng-Hsin, Chirag Rank, Jake Alden Whritner, Olfa Nasraoui, and Patrick Shafto. ‘Unifying Recommendation and Active Learning for Information Filtering and Recommender Systems’. Preprint. PsyArXiv, 25 August 2020. https://doi.org/10.31234/osf.io/jqa83.

Schwalbe, N., & Wahl, B. (2020). Artificial intelligence and the future of global health. The Lancet, 395(10236), 1579–1586. https://doi.org/10.1016/S0140-6736(20)30226-9

Multiple articles from Financial Times - Future of AI and Digital Healthcare

Punn, N. S., Sonbhadra, S. K., & Agarwal, S. (2020). COVID-19 Epidemic Analysis using Machine Learning and Deep Learning Algorithms [Preprint]. Health Informatics. https://doi.org/10.1101/2020.04.08.20057679

This website is the "Resources" archive for the IgniteED Labs at Arizona State University's Mary Lou Fulton Teachers College. The IgniteED Labs allow students, staff, and faculty to explore innovative and emerging learning technology such as virtual reality (VR), artifical intelligence (AI), 3-D printing, and robotics. The left side of this site provides several resources on understanding and effectively using various technologies available in the IgniteED labs. Each resources directs you to external websites, such as product tutorials on Youtube, setup guides, and the products' websites. The right column, "Tech Literacy Resources," contains a variety of guides on how students can effectively and strategically use different technologies. Resources include "how-to" user guides, online academic integrity policies, and technology support services. Rating: 9/10

This is a machine generated response in 2019. We are clearly closer than most people realize to machines that can can pass a text-based Turing Test.

HTM and SDR's - part of how the brain implements intelligence.

"In this first introductory episode of HTM School, Matt Taylor, Numenta's Open Source Flag-Bearer, walks you through the high-level theory of Hierarchical Temporal Memory in less than 15 minutes."

"Amazon's Choice" is made by an algorithm.

Essentially, "Amazon" is Skynet.

I'm so glad I read Dave's post after having just read Rob Horning's great post, "The Sea Was Not a Mask", also addressing algorithms and YouTube.

Great discussion of what questions to ask about artificial intelligence and learning from Dave Cormier.

The new learned optical correlator uses existing light to save energy costs over an optoelectronic two-layer Combines Convolutional neural network (CNN). https://www.sciencedaily.com/releases/2018/08/180802130750.htm

AI and machine learning

Học cơ bản

code: https://github.com/deepmind/dnc

who is curating a list of deep learning success stories, case studies and applications?

such as?

Agree or disagree?

first 15 minutes is a very interesting history of deep learning

2013 article about Douglas Hofstadter, who has continued to pursue an understanding of the human mind through experiments with AI.

https://twitter.com/open_ai

They're hiring: https://openai.com/about/