D.A.D.: Study: Relying on Twitter/X Curated Algorithm Shifts Political Views Rightward — 2/19
The Daily AI Digest
Your daily briefing on AI
February 19, 2026 · 18 items · ~10 min read
From: Cohere, DeepMind, Google AI, Hacker News, Hugging Face Models, Hugging Face Papers, Hugging Face Spaces, OpenAI, Research Blog, arXiv
D.A.D. Joke of the Day
My AI passed the bar exam, medical boards, and the CPA test. I asked it to count the items in my fridge. It said "approximately 12."
What's New
AI developments from the last 24 hours
Anthropic Bars Third-Party Apps From Using Consumer Subscription Logins
Anthropic has published legal documentation stating that OAuth tokens from Free, Pro, and Max subscription plans can only be used with Claude Code and Claude.ai—not third-party tools or services. Developers building products that connect to Claude must use API key authentication through the Claude Console or supported cloud providers. Routing requests through consumer credentials on behalf of users is now explicitly a Terms of Service violation. Community reaction includes frustration over unclear policies across AI companies regarding apps that let users connect their own accounts.
Why it matters: If you're building or using tools that let users authenticate via their Claude subscription rather than paying API costs, that approach is now officially off-limits—clarifying a gray area that affects both indie developers and enterprise teams evaluating third-party Claude integrations.
Deleted Microsoft Guide Allegedly Used Harry Potter for AI Training Examples
A now-removed Microsoft guide allegedly referenced using Harry Potter content for LLM training, with the dataset hosted on Kaggle (which Microsoft owns). The original article has been taken down, leaving only archive links. Community discussion on Hacker News centered on why Warner Bros. or Rowling's representatives haven't pursued legal action, with speculation that the revenue from other Harry Potter ventures makes plain-text infringement a low priority. Others raised questions about whether LLMs trained on copyrighted books could eventually reproduce substantial portions of the original text.
Why it matters: The incident—even scrubbed—highlights unresolved tensions between AI training practices and copyright law, and raises questions about how major tech companies handle potentially infringing content on their own platforms.
Startup Promises Simpler GPU Programming for Rust Developers
VectorWare claims to have implemented Rust's async/await functionality directly on GPUs, allowing developers to use familiar Rust concurrency patterns for GPU programming instead of manual memory and thread management. The company says this could simplify writing high-performance GPU applications. No benchmarks or performance comparisons were provided with the announcement.
Why it matters: This is developer infrastructure—if it works as claimed, it could lower the barrier for Rust developers building GPU-accelerated applications, but practical impact depends on performance data that hasn't been released yet.
What's Innovative
Clever new use cases for AI
Optimized Alibaba Video Model Appears on Hugging Face
A new Hugging Face Space offers what appears to be an optimized version of Alibaba's Wan2.1 video generation model, using technical methods (FP8 precision and AOTInductor compilation) that typically speed up AI model inference. The space is tagged as an MCP server, suggesting it could connect to AI assistants like Claude. No performance benchmarks or documentation were provided.
Why it matters: This is developer infrastructure—potentially useful if you're building video generation workflows and need faster processing, but there's nothing here yet for end users to act on.
Web3-Focused Text Model Released Without Documentation
DMindAI released DMind-3, a text-generation model on Hugging Face tagged for Web3 applications. The model uses what the developers call a 'gpt_oss' architecture and is available in the standard safetensors format. No benchmarks, performance claims, or documentation explaining its capabilities were provided with the release.
Why it matters: This is developer plumbing with minimal information—without benchmarks or clear use cases, there's no way to evaluate whether this offers anything beyond existing open models, even for Web3-focused teams.
Open-Source 'Any-to-Any' Model Launches Under MIT License
A new open-source model called Capybara launched on Hugging Face, claiming to handle "any-to-any" tasks—meaning it can potentially take multiple input types (text, images, audio) and produce multiple output types. Released under the permissive MIT license by xgen-universe, it's built on the diffusers library commonly used for image generation. No benchmarks or performance details were provided with the release.
Why it matters: This is developer plumbing for now—multimodal models that handle diverse inputs and outputs are increasingly common, but without performance data or documentation, there's no way to evaluate whether this offers anything beyond existing options from major labs.
Quiz App Aims to Replace Doom Scrolling—Early Users Are Skeptical
A developer launched Rebrain.gg, a site that replaces passive scrolling with LLM-generated quiz questions—the idea being you learn something instead of mindlessly consuming feeds. Early community reaction has been skeptical. Critics on Hacker News argue the cognitive load is too high to compete with doom scrolling's effortless appeal, with one commenter warning that AI-generated educational content risks letting users 'confidently learn falsehoods.' Others suggested simplifying to swipe-based true/false cards to lower friction.
Why it matters: The skeptical reception highlights a real tension in AI-powered learning tools: making content easy enough to be habit-forming while accurate enough to be educational—a balance that remains unsolved.
Tiny GPT Model Runs Directly in Web Browsers via JavaScript
A Hugging Face Space called 'microgpt.js' has been published by the webml-community, appearing to offer a JavaScript-based implementation of a small GPT model designed to run in web browsers. Details on capabilities and performance are sparse—no benchmarks or documentation accompanied the release.
Why it matters: This is developer plumbing for now: browser-based AI models could eventually enable offline or privacy-preserving AI features in web apps, but this early-stage project isn't ready for business evaluation yet.
What's Controversial
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
Quiet day in what's controversial.
What's in the Lab
New announcements from major AI labs
Major Labs Announce India Initiatives at AI Impact Summit
Google and OpenAI both made major India announcements this week. Google CEO Sundar Pichai announced a $15 billion AI infrastructure investment including a gigawatt-scale computing hub in Visakhapatnam and new subsea cable gateways, calling AI "the biggest platform shift of our lifetimes." OpenAI simultaneously launched "OpenAI for India," an initiative to expand enterprise access through local infrastructure, partnerships, and workforce training. Neither company provided detailed timelines or specific partnership terms. The competing announcements signal India—with its massive English-speaking workforce and fast-growing tech sector—has become a key battleground for AI platform dominance.
Why it matters: The world's most populous country is now seeing direct competition between the two leading AI labs for infrastructure, enterprise relationships, and developer talent—positioning whoever wins to shape AI adoption across emerging markets more broadly.
Gemini App Gets Built-In Music Generation From Text Prompts
Google added Lyria 3, which it calls its most advanced music generation model, to the Gemini app. Users can now create 30-second music tracks from text prompts or images. The feature positions Google alongside competitors like Suno and Udio in the AI music generation space, though Google provided no benchmarks or comparative evidence for its "most advanced" claim.
Why it matters: AI-generated music is moving from standalone apps into major platforms—if you're creating content, presentations, or marketing materials, background music just became a built-in option rather than a separate licensing headache.
74% of Public Servants Use AI, But Only 18% Trust Their Government's Deployment
A survey of 3,335 public servants across 10 countries reveals a striking gap between AI adoption and confidence in government AI strategy. While 74% now use AI in their work and 91% feel confident using it when given clear guidance, only 18% believe their governments deploy AI effectively. The report identifies three tiers: "advanced adopters" (Singapore, Saudi Arabia, India) with strong enthusiasm and training; "uneven adopters" (UK, US, South Africa, Brazil) with inconsistent embedding; and "cautious adopters" (Germany, France, Japan) with limited integration. Top barriers include data security concerns (cited by 50%), unclear organizational rules, and insufficient training. The report—cited by Google at its AI Impact Summit this week—concludes that "ambition, by itself, does not deliver impact."
Why it matters: For organizations rolling out AI internally, this data suggests the gap between employees using AI and believing leadership has a coherent strategy is massive—and that clear guidance, training, and explicit permission structures matter more than executive enthusiasm.
Cohere's Research Lab Releases Lightweight Multilingual Model That Runs on Phones
Cohere Labs released Tiny Aya, an open-weight multilingual model covering 23 languages that's designed to run on consumer hardware including mobile phones. The lab claims it's "the most capable multilingual open-weight model at its scale," with state-of-the-art translation quality. No specific parameter counts, benchmark scores, or comparisons to competing models were provided with the announcement.
Why it matters: For organizations working across languages—especially in markets underserved by English-dominant models—a capable multilingual model that runs locally without cloud infrastructure could reduce both costs and data privacy concerns.
What's in Academe
New papers on AI and its effects from researchers
Study: Relying on Twitter/X Curated Algorithm Shifts Political Views Rightward
A randomized experiment published in Nature assigned nearly 5,000 active US-based X users to either an algorithmic or chronological feed for seven weeks in summer 2023. Users switched to the algorithmic feed shifted toward more conservative policy priorities, were more likely to view criminal investigations into Donald Trump as unacceptable, and adopted more pro-Kremlin views on Ukraine. The algorithm promoted conservative content by 2.9 percentage points, demoted posts from traditional news outlets by 15.5 percentage points, and boosted posts from political activists by 5.9 points. Critically, the effect was asymmetric: switching the algorithm on shifted views rightward, but switching it off did not reverse them — users continued following the conservative activist accounts the algorithm had surfaced. The researchers, from Bocconi University, the University of St. Gallen, and the Paris School of Economics, say this persistent mechanism helps explain why earlier studies on Meta platforms found no political effects from simply turning algorithms off.
Why it matters: This is the first major experimental evidence that X's algorithm doesn't just amplify engagement — it systematically shifts political attitudes in one direction, with effects that persist even after the algorithm is removed, raising fundamental questions about platform power over democratic discourse.
Philosophy Paper Proposes Virtue Ethics as Alternative to AI Alignment Goals
An academic essay argues that AI alignment should abandon goal-based frameworks entirely in favor of virtue ethics. The paper proposes that concepts like 'harmlessness' and 'corrigibility' become brittle when framed as goals or rules, but work naturally when AI systems are designed around 'practices'—networks of actions, dispositions, and evaluation criteria. The formula: instead of 'achieve harmlessness,' design for 'promote harmony harmoniously.' This is theoretical philosophy, not a technical implementation, but it reflects growing interest in alternatives to the dominant reward-maximization paradigm.
Why it matters: As AI companies grapple with alignment failures, this signals that some researchers believe the field's foundational assumptions—not just its methods—may need rethinking.
Humanoid Robots Learn to Grab Everyday Objects Using Plain English Commands
Researchers unveiled HERO, a system that lets humanoid robots manipulate everyday objects—mugs, apples, toys—using natural language commands and vision. The approach combines large vision models with simulated training, achieving a 3.2x reduction in tracking error compared to prior methods. In real-world tests spanning offices to coffee shops, the robot successfully handled objects on surfaces ranging from 43cm to 92cm high. The technical advance: blending classical robotics calculations with machine learning to make humanoid arm movements more precise and adaptable.
Why it matters: This moves humanoid robots closer to practical deployment in warehouses, retail, and service environments where they'd need to handle arbitrary objects on varied surfaces without pre-programming each task.
Best AI Models Still Score 38 Points Below Humans on Spatial Awareness
A new benchmark called SAW-Bench tests whether AI models can understand spatial awareness from first-person video—the kind captured by smart glasses like Ray-Ban Metas. The results expose a major gap: even the best-performing model (Gemini 2.5 Flash) scored 37.66 percentage points below humans on tasks requiring understanding of where the camera is positioned relative to objects. Researchers found models can pick up on partial visual cues but fail to build coherent mental maps of 3D space, causing systematic errors in spatial reasoning.
Why it matters: As tech companies race to build AI assistants for smart glasses and AR devices, this benchmark suggests current models aren't ready to reliably understand 'where am I looking?' and 'what's around me?'—capabilities essential for useful real-world assistants.
AI Agents Still Fail Unpredictably Despite Capability Gains, Study Finds
A research framework proposes twelve metrics across four dimensions—consistency, robustness, predictability, and safety—to evaluate AI agent reliability beyond simple success rates. Testing 14 agentic models across two benchmarks, researchers found that recent capability gains haven't translated to meaningful reliability improvements. Agents showed problems running the same task consistently, handling minor input variations, failing predictably, and containing error severity. The finding suggests that impressive demo performance may not reflect how agents behave in production environments where predictability matters.
Why it matters: For organizations piloting AI agents in workflows, this research validates a common frustration: agents that ace benchmarks can still fail unpredictably in practice—and offers a vocabulary for specifying what 'reliable enough' actually means for your use case.
Masking 0.01% of Training Tokens Boosts AI Reasoning by 7%
Researchers discovered that training instability in AI reinforcement learning stems from a tiny fraction of tokens—about 0.01%—that receive outsized influence during training despite contributing little to actual reasoning. Their fix, called STAPO, simply masks these "spurious tokens" during the learning process. Testing across mathematical reasoning benchmarks with models ranging from 1.7B to 14B parameters, the technique improved performance by an average of 7.13% over existing methods while maintaining more stable training.
Why it matters: This is ML engineering research, but if it holds up, it could make the reinforcement learning phase that improves AI reasoning cheaper and more reliable—eventually meaning better math and logic capabilities in the tools you use.
What's On The Pod
Some new podcast episodes
The Cognitive Revolution — Mathematical Superintelligence: Harmonic's Vlad Tenev & Tudor Achim on IMO Gold & Theories of Everything
AI in Business — Enterprise AI Adoption at a Moment of Maximum Skepticism - with Nishtha Jain