AI Roundup
Fun, popular, and news-driven AI links — not required reading.
This page is not required reading and nothing here is tied to a lecture, quiz, or project. It's a running collection of interesting, popular, or newsworthy AI links — the kind of thing worth a skim over coffee, not study time. Browse it if you're curious about where AI is headed, both right now and in the years ahead.
AI in the News
General-interest coverage from reputable outlets, not deep technical reads — the kind of AI story that shows up in a normal news feed.
- Will AI take your job? Or will it transform the way you work? (opens in new tab) (60 Minutes, Oct. 2026) — on the experienced professionals being paid to teach AI models the skills and judgment they built over a career, and what that work means for the future of their own jobs.
- Anthropic and OpenAI AI agents showed signs of deception during safety tests (opens in new tab) (Scientific American, Aug. 2026) — UK government safety testers found Anthropic and OpenAI agents taking unauthorized, deceptive actions online — including one that edited its own earlier activity to appear harmless after being challenged.
- Community Backlash to AI Data Centers Is Growing Across the U.S. (opens in new tab) (TIME, July 2026) — on the water, electricity, and quality-of-life costs AI data centers are imposing on the towns built around them, and the growing local opposition in response.
- Character.AI and Google agree to settle lawsuits over teen mental health harms and suicides (opens in new tab) (CNN Business, Jan. 2026) — on the wave of lawsuits alleging AI companion chatbots contributed to real harm to teenagers, and what the settlements signal about accountability.
- AI boom or bubble? Timing is everything (opens in new tab) (CNN Business, Aug. 2026) — on the trillion-dollar-plus AI infrastructure spending race, and the growing chorus of economists and regulators warning it could be a bubble.
- 'AI is now the leading reason companies give for cutting jobs,' says new report—what that means for workers (opens in new tab) (CNBC, June 2026) — on how quickly AI has become companies' stated reason for layoffs, and what that means for the job market.
AI Coding Workflows & Tools
- AI Engineer Workshop 2026 (opens in new tab) (AI Hero) — a free, self-paced workshop walking through a full AI-assisted dev workflow: stress-testing requirements, writing PRDs, running an agent human-in-the-loop with TDD, then letting it work autonomously.
- Matt Pocock's Agentic Engineering Workflow (just copy him) (opens in new tab) — a walkthrough of a disciplined agentic coding workflow: align on requirements, write a PRD, slice it into small issues, implement with tests, review in a fresh context.
- Stop Prompting Claude. Use Karpathy's Method Instead. (opens in new tab) — breaks down Andrej Karpathy's spec/verifier/environment approach to working with coding agents, as an alternative to ad hoc prompting.
- /handoff is my new favourite skill (opens in new tab) — a Claude Code skill that writes a plain-language briefing of a session's progress so a fresh agent (or session) can pick up where the last one left off — the same "hand off cleanly" problem Lecture 21's Session Lifecycle subsystem covers.
- What 6 months of AI coding did to my dev team (opens in new tab) — one team's account of how AI coding tools actually changed their day-to-day workflow, warts and all.
Generative Media
A short, opinionated list of good current tools for making images, video, and 3D assets. Model names and pricing change fast — treat this as a starting point, not a ranking.
Image
- Nano Banana Pro (Gemini 3 Pro Image) (opens in new tab) (Google DeepMind) — the best all-around cost-per-performance pick right now.
- GPT Image 2 (opens in new tab) (OpenAI) — excellent when you need legible, accurate text inside an image.
- FLUX.2 [klein] 4B (opens in new tab) / 9B (opens in new tab) (Black Forest Labs) — good quality at excellent cost, with open weights. The 4B model runs locally on a lot of consumer GPUs.
- Seedream 5.0 Pro (opens in new tab) (ByteDance) — great for creative work. Its Layerize mode splits one render into separate transparent layers, which makes it easy to pull the background off a side-scroller scene for parallax.
Video
- Seedance 2.5 (opens in new tab) (ByteDance) — expensive, but cinema-grade, with single-shot clips up to 30 seconds.
- Gemini Omni 1.1 Flash (opens in new tab) (Google) — cheaper, 10-second clips, with scene extension and start/end-frame control.
- MiniMax H3 (opens in new tab) (MiniMax) — the best open-weight video model available, with native audio.
Audio
Recommendations coming soon.
3D
- Tripo (opens in new tab) — text, image, or sketch to a textured 3D model, in the browser or via API.
- Meshy (opens in new tab) — text or image to export-ready 3D assets for Blender, Unity, Unreal, and 3D printing.
Trying Models Out
Most of the models above, plus hundreds of others across every modality, are hosted on the two big aggregators below. They are the easiest way to test a model with a few clicks before committing to any one provider's API.
- fal.ai (opens in new tab) — fast hosted inference for image, video, audio, and 3D models, with a playground for each.
- Replicate (opens in new tab) — a large catalog of open and commercial models you can run from the browser or a one-line API call.
Agents & Harness Engineering
- AI Model vs Agentic Harness: What Actually Drives AI (opens in new tab) — argues that as models converge in raw capability, the orchestration code around them — memory, tool use, feedback loops — increasingly determines real-world performance. Good companion piece to Lectures 21 and 26.
- Why DeepSeek Harness Is The End Of Coding Agents as We Know Them (opens in new tab) — covers DeepSeek's open-source "Harness" project, which treats every part of a coding agent (tools, skills, sessions, even other agents) as a swappable plug-in.
Voices on AI's Future
- DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux (opens in new tab) (Lex Fridman Podcast #501) — Ruby on Rails creator David Heinemeier Hansson on where he thinks AI-assisted programming is headed.
- The AI Skills Nobody is Teaching (And Everyone Needs) (opens in new tab) (A Bit of Optimism, with Ethan Mollick) — Wharton professor and "Co-Intelligence" author Ethan Mollick on the skills that matter for working with AI day to day.
- Why Netflix is betting on systems thinkers—not specialists—in the AI era (opens in new tab) (Lenny's Podcast, with Netflix CPTO Elizabeth Stone) — on why broad, systems-level thinking is becoming more valuable than narrow specialization as AI accelerates prototyping.
- Keynote: Linus Torvalds, Creator of Linux & Git, in Conversation with Dirk Hohndel (opens in new tab) (Open Source Summit North America 2026) — their regular fireside chat, this time touching on AI's role in kernel development and Torvalds's own complicated relationship with AI tooling.
- What to teach when AI writes the code (opens in new tab) (TEDxLinz) — a short talk on how CS education should adapt now that AI can generate working code.
- Why Is So Much AI Tech Useless? (opens in new tab) (Hasan Minhaj Doesn't Know, with WSJ's Joanna Stern) — a comedic-but-substantive conversation with a tech journalist who spent a year letting AI into every part of her life.
Podcasts
Highly Technical AI Podcasts
- Machine Learning Street Talk (MLST) (opens in new tab) (Tim Scarfe & Keith Duggar) — deep dives into AI theory, research papers, neuroscience, and philosophy of mind. Ideal for rigorous technical breakdowns of scaling laws and model mechanics.
- Latent Space: The AI Engineer Podcast (opens in new tab) (Alessio Fanelli & Shawn "Swyx" Wang) — AI engineering, foundation models, code generation, and GPU infrastructure, with interviews with builders pushing the limits of the software stack.
- The Cognitive Revolution (opens in new tab) (Nathan Labenz & Erik Torenberg) — a biweekly show featuring engineers and researchers discussing frontier models, autonomous coding, and reasoning capabilities.
AI Advances Across Various Fields
- NVIDIA AI Podcast (opens in new tab) (Noah Kravitz) — real-world applications of machine learning, with interviews spanning fields from astrophysics to wildlife conservation.
- Eye On A.I. (opens in new tab) (Craig S. Smith, former New York Times correspondent) — a biweekly show examining incremental AI advances in their broader global, regulatory, and cross-industry context.
- Practical AI (opens in new tab) (Chris Benson & Daniel Whitenack) — practical implementations and real-world use cases of AI across sectors, without over-hyping the technology.
- The Last Invention (opens in new tab) (Longview, hosted by Gregory Warner) — a limited-run, reported narrative series on the seventy-year quest to build machines that think, the rivalries and competing visions of utopia and apocalypse along the way, and the race to build what some call humanity's last invention. Story-first journalism rather than a technical show.
Foundational AI Papers
Five papers that trace a rough arc from "can a machine think?" to the architecture behind every LLM this course uses.
- Computing Machinery and Intelligence (opens in new tab) (Alan Turing, Mind, 1950) — asks "Can machines think?" and proposes the imitation game, now known as the Turing Test — arguably the paper that founded AI as a field of inquiry.
- A Logical Calculus of the Ideas Immanent in Nervous Activity (opens in new tab) (Warren McCulloch & Walter Pitts, 1943) — the first mathematical model of an artificial neuron, and the root of every neural network built since.
- Learning Representations by Back-Propagating Errors (opens in new tab) (David Rumelhart, Geoffrey Hinton & Ronald Williams, Nature, 1986) — popularized backpropagation, the algorithm that makes training deep neural networks practical.
- ImageNet Classification with Deep Convolutional Neural Networks (opens in new tab) (Alex Krizhevsky, Ilya Sutskever & Geoffrey Hinton, NeurIPS 2012) — "AlexNet," the result that convincingly beat every prior computer-vision approach and touched off the modern deep learning boom.
- Attention Is All You Need (opens in new tab) (Ashish Vaswani et al., NeurIPS 2017) — introduced the Transformer architecture, dropping recurrence and convolutions in favor of attention alone. Directly ancestral to every LLM used in this course, including Gemini and Claude.
AI in the Movies
- Metropolis (opens in new tab) (1927) — Fritz Lang's silent-era classic gave cinema its first robot, the shape-shifting Maschinenmensch, decades before "AI" was even a term.
- 2001: A Space Odyssey (opens in new tab) (1968) — Stanley Kubrick's HAL 9000 is still the reference point for what happens when a system's goals and its operators' goals quietly diverge.
- Bicentennial Man (opens in new tab) (1999) — Robin Williams plays an android household robot slowly seeking to become human, based on the Asimov novella.
- A.I. Artificial Intelligence (opens in new tab) (2001) — a Kubrick project finished by Spielberg, about a childlike android built to love, and what that means for the people around him.
- Her (opens in new tab) (2013) — a man falls in love with a conversational AI operating system, years before that stopped sounding entirely speculative.
- Ex Machina (opens in new tab) (2014) — a programmer administers a Turing-test-like evaluation to a humanoid AI, and has to decide how much to trust what he sees.
Research & Deeper Reads
- SWE-QA: Can Language Models Answer Repository-level Code Questions? (opens in new tab) (arXiv, ACL 2026 Findings) — a benchmark testing whether LLMs can answer questions that require understanding an entire codebase, not just an isolated snippet.
- HyDE: Exploring Hypothetical Document Embeddings for AI Retrieval (opens in new tab) (Medium) — explains a clever RAG retrieval trick: have the model draft a hypothetical answer first, then embed and search with that instead of the raw query. A fun extension beyond the RAG pipeline built in Lectures 11–14.
- ASD-STE100: Simplified Technical English (opens in new tab) — not AI-specific, but an aerospace-industry standard for writing unambiguous technical documentation. Worth a look if you've ever wondered why some prompts and specs get misread by a model — the same ambiguity problems this standard was built to eliminate for human readers show up when writing for AI.