AI

April 2026

• Apr 10, 2026Work SkillsSoftware Engineering Career Outlook in 2026Amazon cut roughly 18,000 roles in January 2023 while still hiring for AWS infrastructure and AI, and Google shed 12,000 while expanding DeepMind and cloud teams, a pattern the page calls portfolio rebalancing rather than falling demand. Over 350,000 tech layoffs ran from January 2022 to December 2024, yet the Bureau of Labor Statistics projects 26% growth in software developer employment through 2032.17 MIN7.9k
• Apr 9, 2026Work SkillsIs Product Management a Good Career in 2026?Google's APM program draws more than 10,000 applications for roughly 50 spots, and LinkedIn PM openings have recovered to about 80% of their mid-2021 peak. The page argues the 2022 to 2024 cuts were corrections toward pre-pandemic headcount ratios rather than structural decline, since many firms had hired two to three times sustainable PM levels. The hard market is for juniors and generalists, not specialists.19 MIN11.3k
• Apr 6, 2026Work SkillsData Science Career Value in 2026Indeed and LinkedIn listings grew about 28% a year from 2019 to 2022, fell roughly 35 to 40% from the Q3 2021 peak through late 2023, then recovered to sit around 20 to 25% below that peak. That leaves a market still larger than 2018 but far more demanding. The page's answer is explicitly conditional, depending on which tier you enter and which specialization you build, since the field sorted itself out rather than collapsing.16 MIN2.4k

January 2026

• Jan 14, 2026TechnologyLarge Language Models ExplainedSelf-attention lets a model weigh every word against every other word at once, and next-word prediction at trillion-token scale forces grammar, world knowledge and reasoning patterns into the weights. The article is candid that hallucination follows from generating for likelihood rather than truth. Hoffmann's Chinchilla work showed a 70-billion-parameter model trained on 1.4 trillion tokens beating GPT-3's 175 billion on most benchmarks.21 MIN56.1k

September 2025

• Sep 18, 2025IdeasAI Use Cases for Small TeamsBrynjolfsson, Li and Raymond studied 5,179 customer support agents and also recorded 9 percent higher satisfaction scores, with the largest gains going to newer workers who effectively reached expertise that normally takes years to build. For a team that cannot afford specialists, that makes these tools a knowledge multiplier. Dell'Acqua, Lakhani and colleagues found a similar floor-raising effect across 758 consultants.31 MIN18.4k

June 2025

• Jun 9, 2025ExplainersAn Overview of Machine Learning Model TrainingKrizhevsky, Sutskever and Hinton won using ReLU activations, dropout and GPU acceleration, techniques that were already available but had not been combined this way. That is the real lesson of 2012. This page works through the training loop from first principles: loss functions, gradient descent and the Adam optimizer, backpropagation and vanishing gradients, overfitting and regularization, why data is split three ways, and how to tell when training is finished.45 MIN58.5k

June 2024

• Jun 9, 2024ExplainersHow AI Learns: Neural Networks, Training, and the Math Behind MLHand-coded rules work for chess and fail for faces, so the field moved to adjusting millions of parameters against an error signal. The page dates the turn precisely to 2012, when a deep network cut the ImageNet error rate from about 26 percent to 15 and computer vision pivoted almost immediately, and it is honest about the two earlier winters when the same idea stalled for want of data and compute.19 MIN6.8k
• Jun 4, 2024CultureThe Future of Work: Remote, AI, and Automation InsightsOECD researchers re-ran the famous Oxford analysis at the task level and found 9 percent of US jobs at high risk rather than the headline figure, which the page itself calls overstated. A nurse's work spans data entry and comforting dying patients, so the job survives while pieces of it are eaten. Acemoglu's robot study adds the harder point, that who collects the gains depends on institutions, not the technology.15 MIN11.8k

August 2023

• Aug 26, 2023ExplainersThe Inner Workings of Machine Learning ExplainedHow machine learning works, from training data through backpropagation to why models overfit. In the 2012 ImageNet challenge, AlexNet reached a top-5 error rate of 15.3 percent against 26.2 percent for the runner-up, a gap that launched modern deep learning. Training data decides the outcome: the 2018 Gender Shades study found commercial gender classification erred on dark-skinned women as often as 34.7 percent of the time.21 MIN14.1k

May 2023

• May 17, 2023Work SkillsAI Transformations in Software Engineering: Key FindingsWhat the research actually shows about AI and software engineering. A 2025 METR study found experienced open-source developers were 19 percent slower using AI tools on familiar codebases, likely because reviewing, prompting and correcting output offset the generation time saved. GitHub's own controlled study found a 55 percent speed gain, but on a small well-specified task that favors language models. The measured gain depends entirely on what the engineer is doing.17 MIN2.5k
• May 15, 2023Work SkillsAI's Influence on Data Science: Key Changes AheadWhat generative AI actually changed in data science work. Text classification that once meant TF-IDF plus a trained classifier over days of work is now prompt engineering and an API call measured in hours, and RAG pipelines have replaced search-plus-manual-reading for document questions. AutoML genuinely speeds up baselines on clean tabular data, but it cannot frame the problem, will not detect data leakage or biased sampling, and handles text, images and graphs poorly.15 MIN2.3k

October 2022

• Oct 29, 2022TechnologyWhat Is Prompt Engineering: Improving AI Output QualityFor most of the history of natural language processing, a new task meant labelled examples, fine-tuned weights and a separate model per job. Large language models moved that work into the instruction, so what you adjust is the prompt rather than the weights. The core techniques are zero-shot requests, few-shot examples, and chain-of-thought, which Wei and colleagues found in some cases more than doubled accuracy on multi-step reasoning.16 MIN10.6k

September 2022

• Sep 12, 2022TechnologyWhat Is Transfer Learning in AI: How Models Reuse KnowledgeYosinski and colleagues showed at NeurIPS 2014 why it works: the first layers of a network trained on ImageNet are essentially universal, middle layers transfer moderately, and only the final layers are strongly task-specific. That layer-by-layer result underpins everything from fine-tuning a ResNet on a few thousand medical scans to LoRA, which injects low-rank matrices into a frozen model and cuts trainable parameters by a factor of 10,000 while keeping most of the gain.21 MIN3.1k