Click any tag below to further narrow down your results
Links
After Google told schools they could turn off Gemini but not its AI-laden Search, DuckDuckGo is pitching itself as a privacy-focused, AI-optional alternative to win education customers. The company says it offers its own AI features but lets users disable them everywhere, unlike Google.
- Google said schools can disable Gemini, but turning off its other AI features would also cut off Google Search, which now relies heavily on AI Overviews and AI Mode.
- Google's Chromebooks are in roughly 50 million schools, and Google Classroom serves over 150 million users, so the education market is a major stake for Google.
- DuckDuckGo says its app installs rose 30% earlier this year as users rejected AI-forced Google Search, and it now directs schools to switch their default search engine.
Google is reportedly planning to restrict or turn off its search and other products for schools that refuse to adopt its A.I. tools, a move that would give the company leverage over classrooms it already dominates. The piece, by Dana Goldstein and Natasha Singer, looks at what that pressure means for districts that want to keep AI out of their students' hands.
- Google products are already embedded across American schools, which makes a cutoff a serious threat for districts that depend on them.
- The policy would tie access to Google's search and related services to schools' willingness to use its A.I. offerings.
- The article raises concerns about how much control a single company gains over what students and teachers use in class.
humans& has released &Tutor, a model trained to teach by getting rewarded for how much a simulated learner (Persimmon) picks up. It scores 0.754 on the TutorMoments benchmark against 0.500 for its base model and beats several frontier models, and the team says the learner model's realism is what makes the tutor work.
- On TutorMoments under a plain prompt, &Tutor scores 0.754, versus 0.500 for the GLM-5.3 base model and 0.667 for Claude Sonnet 5.5.
- Teaching behaviors like knowing when to scaffold or push for rigor are never trained directly; they emerge from rewarding learning gains.
- Swapping in GLM-5.3 as the learner instead of Persimmon produced a noticeably weaker tutor, which the authors attribute to Persimmon being a more realistic learner.
The author argues that programmers shouldn't dismiss AI code generation as harmless just because programming is "functional" rather than "artistic" — programming is art, and using AI to eliminate mundane work destroys the practice opportunities where developers actually learn their craft. Most meaningful creative work in any field starts with doing unglamorous, paid work, and automating away those jobs prevents the skill development that leads to genuine innovation.
- Art isn't defined by being transcendent or non-functional; most painting, music, and writing is mundane, commissioned work that serves practical purposes, yet we still call it art and respect the labor involved. Programmers are uniquely willing to dismiss their own work as "just functional" in ways writers, musicians, and artists refuse to do.
- Code can be intentionally beautiful and aesthetically designed (like the author's Deferred abstraction for async operations), and such work tends to be socially and technologically important — but you only develop that capability through years of doing unglamorous projects that build your skill and intuition.
- Replacing all mundane programming work with AI-generated code removes the practical training ground where most real skill development happens, eliminating the tiny but real chance that a routine project becomes someone's breakthrough moment and career-defining contribution.
Mathai is an AI service that breaks down math problems into line-by-line solutions quickly. It handles algebra, calculus, graphs, and equations—useful for students who need answers fast or want to check their work.
- Mathai gives step-by-step breakdowns of math problems (algebra, calculus, graphs, equations) instead of just final answers, aimed at speed for time-crunched students
- It's positioned explicitly as a utility for checking work or getting unstuck, not as a tutor focused on learning
- The article flags an unaddressed gray area: using it to rush assignments or verify answers could clash with school academic honesty policies
This is an AI tool that solves chemistry problems by breaking them into structured steps: setup, units, calculations, and final answer. It's designed for students who need quick homework verification or last-minute practice, with built-in unit conversion handling for common chemistry calculations like moles and molarity.
- Chemistry AI breaks problem solutions into setup, unit conversions, calculation steps, and final answer.
- It's aimed at exam cramming, homework verification, and lab-related stoichiometry/molarity calculations.
- The tool enforces unit safety to help catch calculation errors that produce nonsensical unit combinations.
- The page gives no details on pricing or which chemistry topics (e.g., organic chemistry, thermodynamics) are actually covered.
Physics AI is an AI tool that solves physics problems in two ways: Solver Mode gives you instant answers with step-by-step calculations for quick checking, while Tutor Mode uses the Socratic method to guide you through concepts. It handles everything from mechanics to quantum mechanics, generates diagrams and free-body sketches, and works with both typed problems and photos of handwritten work.
- Offers two distinct modes: Solver Mode for instant step-by-step answers, Tutor Mode using Socratic questioning to build understanding
- Generates visual aids like free-body diagrams, pulley/optical sketches, and kinematic/thermodynamic graphs, not just equations
- Accepts photos of handwritten problems or diagrams alongside typed input
- Free tier gives 7 credits/month; paid plans run $5.90–$11.90/month for 500–1,200 credits, with 40% off annually
@not_ellington on X
A post from @not_ellington argues that sitting down to read and reread primary material beats the viral "educational" threads that rack up thousands of bookmarks. It links to an arXiv paper (2606.24937) as the kind of serious source worth that effort.
- The author calls viral 15-minute learning threads "eduslop," saying they get heavy bookmarking but teach less than the source material they summarize.
- The core claim is that the best way to learn is still to read something carefully and then reread it.
- The linked arXiv paper is offered as an example of the depth the author means, rather than a summary of a thread.
@ClarkSquareCap on X
The post recommends Joel Greenblatt's free lecture notes from his Columbia "Special Situation Investing" course as one of the best free investing resources online. The author points to Greenblatt's roughly 40% annual return over ten years and says the course is the one he wished he'd had as a student.
- The notes come from Joel Greenblatt's Columbia class on special situation investing, shared as a free PDF.
- The author credits Greenblatt with one of the best 10-year track records in investing, around 40% a year.
- The course is pitched as the class Greenblatt wished he'd had as a student, which is the author's main reason for calling it a top free resource.
A hacking group breached the University of Nottingham’s PeopleSoft student records system and stole over 40 GB of data on 454,600 current and former students, including names, addresses, financial details and academic records. The university has reported the incident to the UK Information Commissioner’s Office and Action Fraud, while ShinyHunters claims responsibility and posted the stolen archive.
- ShinyHunters stole 40GB of data on 454,600 current and former Nottingham students, including financial details, credit card info, passport numbers, and academic records.
- The group exploited zero-day and older vulnerabilities in Oracle PeopleSoft, hitting over 100 organizations in this campaign, with Oracle yet to confirm active exploitation.
- Similar PeopleSoft-linked breaches recently hit Oxford University's CareerConnect and Instructure's Canvas LMS.
@EmoPhilips on X
The post from @emophilips shares a catchy mnemonic designed to help people memorize the alphabet. Beyond that single idea, the content is thin, so there's little else to dig into.
- A memorable mnemonic is offered as a way to learn the alphabet.
This paper defines the “LLM fallacy” as a bias where users credit their AI-assisted outputs to their own skill rather than the model’s contribution. It analyzes how fluent, opaque interactions with large language models blur human-machine boundaries, offers a framework for its mechanisms, and discusses impacts on education, hiring, and AI literacy.
- People using LLMs tend to internalize the AI's output as evidence of their own skill, not just trust the tool's answer (unlike classic automation bias).
- This happens because model reasoning is invisible, the prose feels fluently "authored," there's no clear handoff moment, and positive feedback loops reinforce the false belief.
- It shows up concretely: coding, academic writing, data analysis, and creative work all get passed off as personal competence when AI did much of the work.
- Real consequences include students graduating without real mastery, hiring managers misjudging AI-reliant candidates, and skewed performance reviews at work.
This article breaks down Andrej Karpathy’s zero-dependency, 243-line GPT implementation in plain Python. It explains how each part—tokenizer, autograd engine, embeddings, attention mechanism, residual connections, and MLP—mirrors a full-scale transformer on a tiny dataset of baby names.
- Karpathy's microGPT implements a full GPT—tokenizer, autograd engine, transformer, training loop—in just 243 lines of pure Python with zero external dependencies beyond os, math, random and argparse.
- A ~40-line custom autograd engine (Value class) replicates PyTorch's backward-pass mechanics via topological graph traversal.
- The toy model trains on baby names using a tiny architecture (16-dim embeddings, seq length 8, single layer, 4 attention heads) totaling about 4,000 parameters.
- The same core transformer math—embeddings, RMSNorm, attention, MLP—scales up unchanged to power trillion-parameter models like GPT-4.
@RohOnChain on X
The post promotes a two-hour Stanford lecture on building an AI career, and the author argues it beats most of the AI commentary people scroll past. The post itself is a short endorsement with no claims beyond the lecture's length and the author's recommendation.
- A two-hour Stanford lecture focused on AI careers is the recommended content.
- The author claims it teaches more about winning in the AI race than the usual AI content online.
- The author urges readers to set aside the full two hours this weekend.