weekend ai reads for 2025-12-12

For now, this task stands undefeated. A monument to 1996 web design and a humbling reminder that sometimes the simplest tasks are the hardest. That orbital pattern of planets, thrown together by some Warner Brothers webmaster 28 years ago, has become an inadvertent benchmark for Claude.

Until then, the Space Jam website remains proof that not everything old is obsolete. Some things are just irreproducibly perfect.

Only Gemini 3 solved this riddle. Even Opus 4.5 couldn't. And Grok? Good lord... someone check on Grok.

It shows how brittle these intelligences are. To see their respective failure modes (and Gemini 3’s impressive performance) read on.

Move slowly and methodically to go fast

It seems fairly safe to assume you aren’t in the first group against the wall, if you’re excited about the thing putting them there.

Crucially, AI is not just helping people do the same work faster—it is enabling people to do new kinds of work.

Coding-related messages increased 36% for workers outside of technical functions.

75% of users report being able to complete new tasks they previously could not perform.

AI reduces the gap between intent and execution. Individuals can translate ideas into concrete outputs, regardless of their specialization or technical expertise.

The promise of AI – the promise AI companies make to investors – is that there will be AIs that can do your job, and when your boss fires you and replaces you with AI, he will keep half of your salary for himself, and give the other half to the AI company.

Clients can use LLMs to compress my advice into summaries without telling me. Pressure to keep writing short and plain is good, but they’re losing my careful choice of how much to generalize rules and realities. They’re also interrupting how I choose to teach and reinforce useful vocabulary.

Nano-Banana Pro is a “Thinking” model. It doesn’t just match keywords; it understands intent, physics, and composition. To get the best results, stop using “tag soups” (e.g., dog, park, 4k, realistic) and start acting like a Creative Director.

Netflix feels like a comfortable sofa. You sink into it. The other streaming services feel like you’re shopping.

This is the moat that AI companies keep ignoring. They’re obsessed with model capabilities—who has the best reasoning, who generates the most accurate images, whose context window is bigger. Meanwhile, the real competition is happening at the interface level, and most of them are losing.

Across studies, we find that genAI-created ads consistently outperform both human-and genAI-modified ads, increasing click-through rates by up to 19% in field settings. In contrast, genAI-modified ads show no significant improvement over human-created benchmarks. These results reveal an asymmetry: visual genAI delivers greater value when used for holistic ad creation rather than for modification, where creative constraints may limit its effectiveness. … Finally, we find that disclosing AI involvement in ad generation significantly reduces advertising effectiveness by up to 31.5%, underscoring trade-offs relevant to evolving AI disclosure policies.

So Estonia is pioneering a better way of using ChatGPT to learn — by making it harder and retraining the model to pose more questions than answers. The effort challenges ChatGPT’s core appeal and could get the cold shoulder from OpenAI, but it could also unlock AI’s true potential in education in a way no one else has managed.

Being reared by robots has advantages. Tech firms are already showing how AI can enhance learning, especially where teachers and materials are scarce. Literacy and language-learning have been boosted in early trials. The dream is that, with an AI tutor, children can be saved from classes pitched to the median, in which bright pupils are bored and dim ones are lost.

Rather, we seek to determine whether existing models—which can be straightforwardly applied by most professors and students—are already suitable for the task of law exam evaluation. We find that, when provided with a detailed rubric, the LLM grades correlate with the human grader at Pearson correlation coefficients of up to 0.93. Our findings suggest that, even if they do not fully replace humans in the near future, LLMs could soon be put to valuable tasks by law school professors, such as reviewing and validating professor grading, providing substantive feedback on ungraded midterms, and providing students feedback on self-administered practice exams.

The AAIF provides a neutral, open foundation to ensure this critical capability evolves transparently, collaboratively, and in ways that advance the adoption of leading open source AI projects. Its inaugural projects, AGENTS.md, goose and MCP, lay the groundwork for a shared ecosystem of tools, standards, and community-driven innovation.

We investigate why organizations build agents, how they build them, how they evaluate them, and what the top development challenges are. We find that production agents are typically built using simple, controllable approaches: 68% execute at most 10 steps before requiring human intervention, 70% rely on prompting off-the-shelf models instead of weight tuning, and 74% depend primarily on human evaluation. Reliability remains the top development challenge, driven by difficulties in ensuring and evaluating agent correctness.

Grok ended up performing the best while DeepSeek came close second. Almost all the models had a tech-heavy portfolio which led them to do well. Gemini ended up in last place since it was the only one that had a large portfolio of non-tech stocks.

Similar to Coke’s 2025 Holiday ad, the McDonald’s spot is like a visual seizure, full of grotesque characters, horrible color grading, and hackneyed AI approximations of basic physics.

Though the abomination of an ad only has 20,000 views on YouTube, backlash in the comments was so intense that McDonald’s shut down comments over the weekend, before delisting the video entirely. (Some marketing research databases managed to scrape the clip, if you’re curious.)