EP01 - Saturday Nights with Technology : Saturday, September 12, 2026

Hello, and welcome to Saturday Nights with Technology, I am Zarif Ahmed. This week we have three significant technology stories. Apple has finally entered the foldable phone market with iPhone Duo. Anthropic has accused several Chinese AI companies of improperly using Claude to help train competing models. And two major AI systems, Claude Table 5.1 and OpenAI's GPT-6 Astra, have arrived with a stronger focus on professional and agentic work. Later in the program, I will also review the book AI 2041 and introduce two open-source GitHub projects worth exploring OpenMIC and Cogivolve4B. Let's begin with Apple. After years of watching Samsung, Huawei, and other manufacturers experimenting with foldable phones, Apple has finally officially joined the category with the iPhone Duo. The name Duo reflects the central idea behind the device. It can operate like a conventional phone when folded, but open into a larger tablet-like screen when open. The iPhone Duo has a 5.4-inch external display and a 7.6-inch internal folding display. Apple has built it around the A20 Pro processor, a titanium body, and a newly designed e-System. The company says the hardware and iOS 27 have been designed together to support applications across both screen sizes. Apple is positioning the larger display as a productivity feature rather than merely a bigger screen. Users can work with multiple applications, view more information, and move between phone and tablet-style experiences without carrying a second device. There are, however, some interesting compromises. Apple has reportedly replaced Face ID with Touch ID built into the side button. And the device also removes the physical SIM tray in supported markets and does not include the familiar action button. The most striking detail may be the price. The iPhone Duo starts at $1,999 in the United States and approximately 2,99,900 rupees in India. Pre-orders are scheduled to begin on October 16, with availability starting on October 23. That price places the Duo well beyond the conventional premium phone category. It is closer to the combined cost of a high-end phone and a capable laptop. The most important question is not whether Apple can manufacture a polished tablet phone or foldable phone, it almost certainly can. The real question is whether the larger screen creates enough practical value to justify the price, particularly when many buy us already on a phone, a tablet, and a laptop like me. Apple is entering the foldable market relatively late, but Apple has been often preferred to enter an established category after the technology has matured, then attempt to make it attractive to a much larger audience. Whether the iPhone Duo can do well will depend on three things, the durability of its folding display, the quality of the application designed for the new format, and perhaps most importantly, whether ordinary users discover a convincing reason to unfold their phone. Our second story concerns a growing dispute over how artificial intelligence models learn from one another. Anthropic has accused Chinese AI companies including DeepSeek, Moonshot, AI, and Minimax of conducting what it describes as an industrial-scale campaign to extract capabilities from clouds. The technique at the center of all the dispute is called model distillation. In legitimate distillation, a larger and more capable model generates answers that are used to train smaller models. This can make the smaller model faster, cheaper, and more efficient. Distillation itself is a recognized machine-learning technique and is not automatically improper. The controversy begins when a company repeatedly queries a computer's commercial model, collecting the responses, and using those outputs to reproduce its capabilities without permission. Anthropic alleges that the company used large numbers of accounts and intermediary services to send requests to clouds while disguising who was actually making them. According to Anthropic, the purpose was not simply to use cloud as a customer but to obtain enough structured output to improve competing AI models. These remain allegations, and this should be described that way unless independently established or accepted by companies involved. The story also exposes an uncomfortable issue for the entire AI industry. AI companies have trained their own systems using enormous amounts of material produced by writers, programmers, artists, publishers, and websites, and some of the material was used without direct negotiation with its creator. Even just last week, on one of my websites, I ended up paying $200 for the bandwidth only because Cloude was crawling at a higher speed. Now, these same companies are trying to prevent competitions from using the output of their models for training. Legally, the question may be different. Commercial terms of services can prohibit automated extraction, credential misuse, or training competing systems from model response, but ethically, the distinction will continue to be debated. Where do ordinary learning and model theft begin? How many responses must be collected before normal use becomes systematic extraction? And can companies demand strict protection for model outputs while maintaining that training on publicly available human work is acceptable? This is no longer merely a technical disagreement. It is becoming a geopolitical and commercial conflict over who owns artificial intelligence capabilities and how those capabilities may be transferred. That brings us to our third story, the arrival of Cloud Fable 5.1 from Anthropic and GPT-6 Extra from OpenAI. Cloud Fable 5.1 is positioned as Anthropic's most capable model for demanding easy coding, research, and long-running agentic tasks. The phrase agentic tasks is important here. These newer models are not designed only to answer a single question. They are increasingly expected to examine files, use software, tools, browse information, write and test code, identify mistakes, and continue working through tasks over an extended period. Anthropic says Fable 5.1 improves upon the previous Fable models in areas such as coding and professional knowledge work. The company has also released Cloud Mythos 5.1 as part of the same generation. OpenAI, meanwhile, has introduced GPT-6 Extra. Extra is being presented as a major improvement in compute use, browsing software, engineering cybersecurity, scientific reasoning, and professional work. It is available through Chet-GPT Work, Codex, and the API. Although, access may continue to expand gradually across different accounts and services. One of the major, most interesting aspects of Extra is the emphasis on computing work rather than simply producing answers. It can interact with computer interfaces, work across multiple steps, and create documents, spreadsheets, and presentations while responding to changing instructions. This reflects a wider shift in the AI industry. The first phase of Generative AI was largely about conversations. We asked a question and received an answer. The next phase was about creation. Models wrote articles, generate images, and produce software code. The emerging phase is about execution. The model is expected to understand a goal, decide which tools to use, perform a sequence of actions, verify the results, and recover when something goes wrong. That sounds powerful, but it also introduces new risks. A chatbot that produces an incorrect paragraph may be a inconvenience to someone. An autonomous system that changes production codes and information, operates business softwares, or modifies infrastructure can cause many problems. As these models become more capable, evaluation must move beyond benchmark scores. Organizations will need to ask practical questions. Can the model recognize when it lacks sufficient information? Can administrators see what it has done? Can its access be restricted? And can a human reverse its actions when it makes a mistake? For developers and technology teams, the competition between Fabel 5.1 and GPT-6 Extra will not be settled by one benchmark table. The better model will depend on the work being performed, the tools surrounding it, the quality of reasoning, its reliability over long tasks, its security controls, and the total cost of completing the job. Now it is time for this week's book review. The book is AI 2041. Those who are on the video can see it is in my hand. Those who are listening on the audio only mode. So this book was published in 2021 before the arrival of chat GPT turned generative AI into an everyday public technology. The timing makes the book especially interesting to revisit today. AI 2041 uses an unusual but effective structure. It contains 10 fictional stories set in different parts of the world, with each story followed by a factual explanation of the technology behind it. Writer gives us the human story. Then it follows with the explanation, the artificial intelligence concept, and the other considerations about how realistic that image in the future may be. The subject includes personalized education, deep fakes, autonomous vehicles, health care, job displacement, immersive digital experience, privacy, and the possible economic abundance created by automation. The book's biggest strength is its accessibility. It does not expect the reader to understand neural networks, training methods, or complex mathematics. Instead, it begins with the pupil and the situation, and then introduces the technology through its consequences. This approach also makes the book useful for business leaders, teachers, and policymakers. It encourages them to think beyond what an AI model can do today and consider how its use might change an organization or an entire society. Another strength is its international setting. The future is not imagined only for Silicon Valley. The story takes place across countries and cultures, showing that AI will not affect every community in exactly the same way. However, the book also has limitations. Its overall attitude towards AI is overly optimistic, sometimes more optimistic than recent experience may justify. The technical possibilities are explored in detail, but the question of the concentrated corporate power, unequal access, and political control of AI could have received more attention. Some predictions also feel surprisingly different. The progress of generative AI since 2001 has been much faster in certain areas than many people expected. But that does not make the book outdated. In fact, it creates an additional reason to read it. I can compare its image in 2041 with what has already happened and ask whether the remaining change could also drive earlier than expected. My recommendation is that AI 2041 is worth reading if you want a thoughtful introduction to the social consequences of artificial intelligence without reading a heavy technical book. It is not a manual for building an AI system, by the way, and it is not a perfect prediction for the future. It is a collection of scenarios designed to make us ask better questions about the future we are already beginning to build. Let's finish our recommendation with two connected open source projects. The first repository is OpenMIC on GitHub. And OpenMIC stands for Open Multi-Agent Interactive Classroom. It takes a topic or an uploaded document and turns it into an interactive learning experience. Instead of giving the learner another chatbot window, it creates something closer to a classroom. AI teachers and AI classmates can present lessons, conduct discussions, ask questions, and use a shared whiteboard. The platform can generate slides, quizzes, interactive HTML simulations, and project-based learning activities. It also supports text-to-speech, speech recognition, and export of the editable PowerPoint presentation and interactive HTML pages. I have even used some TTS software to create my own voice, and it can generate the tutorial in my own voice, actually. By the way, it is me, Riyal. And what makes it wonderful and interesting is that you can run it in your own machine, if you have that kind of, you know, power. And you can try a hosted version also, by the way, or you can self-host it and connect it to several model providers, including local models through OLAMA. So I have used OLAMA, I have used OpenAI also for this. And for trainers and educators, this repository is worth studying, even if they do not deploy it directly. It demonstrates how AI could change digital learning from passive content delivery into a more participatory experience. The second repository is CogEvol 4B. CogEvol 4B is a compact, specialized model designed to turn a short post description into a complete learning artifact, such as structured slides or self-contained interactive HTML page. Its quantized model is approximately 2.4 GB and can run locally through OLAMA. According to its developer, it can operate on Apple Silicon, I have tried it on Apple Silicon only, CUDA, CUDA hardware, and even on a CPU without requiring a cloud API key during the generation. It also integrates with OpenMAC. That means the two repositories can be explored as part of the same stack. OpenMAC provides the classroom experience, while CogEvol 4B provides a specialized local model for generating learning material. So instead of connecting to OLAMA or OpenAI or cloud AI, you can just use this 4B model. So they are still practical considerations. Generated lessons may be reviewed for accuracy. They should be. Interactive code should be treated as untrusted, and local performance will depend on the available hardware. On my Mac Silicon, it was faring pretty much fine. So that is all for this week. Taken together, this week's stories point towards the same larger development. Our devices are changing shapes, AI companies are fighting over how intelligence is acquired, and AI models are moving from answering questions to performing real work. The iPhoneView asks whether a phone can become a practical second screen. Anthropic allegations ask who owns the knowledge represented inside an AI model. And the release of Cloud Fable 5.1 and GPT-6 Extra asks how much responsibility we are prepared to give these systems. AI2041, OpenMAC, and CogEvol 4B then bring the discussion back to the pupil. How we imagine the future. How we learn the weather. Powerful AI experience can be built locally and openly. And the future of technology will not be determined by the capability alone. It will also depend on prices, access trust, ownership, and the control. That's all for this technology update. I'm Zarif Ahmed, and you have been listening to Saturday Night with Technology. Next week, I'll come back with another episode in which I'll just summarize the week, will introduce you to maybe another book, and maybe some better available GitHub repositories that you can try on your local machine. Till then, bye. Thank you.