D.A.D.: Humans Crush Even The Best AIs At Video Games — 2/20
The Daily AI Digest
Your daily briefing on AI
February 20, 2026 · 13 items · ~8 min read
From: DeepMind, Google AI, Hacker News, Hugging Face Models, OpenAI, arXiv
D.A.D. Joke of the Day
My AI gave me three versions of every email I asked for. Now I spend more time choosing than I ever spent writing.
What's New
AI developments from the last 24 hours
Google's Gemini 3.1 Pro Posts Top Scores Across Most Major Benchmarks
Google released Gemini 3.1 Pro with benchmark results that, if verified, put it at or near the top of the field on most major tests. On ARC-AGI-2, which measures novel reasoning, 3.1 Pro scored 77.1%—ahead of Opus 4.6 (68.8%), Sonnet 4.6 (58.3%), and GPT-5.2 (52.9%). It also leads on Humanity's Last Exam (44.4%), LiveCodeBench Pro (2887 Elo vs GPT-5.2's 2393), and competitive coding benchmarks like SciCode (59% vs Opus 4.6's 52%). It falls short on SWE-Bench Verified, where Opus 4.6 edges it (80.8% vs 80.6%), and on SWE-Bench Pro, where GPT-5.3-Codex leads (56.8% vs 54.2%). The model is rolling out across Google's developer tools, enterprise products, and consumer apps. Early community reaction on Hacker News has been skeptical, with some arguing the version jump from 3.0 to 3.1 overstates what they expect is a modest upgrade.
Why it matters: The stated benchmarks represent a genuine leaderboard shakeup—Google matching or beating Anthropic and OpenAI's best on most major reasoning and coding tests. The skepticism is worth noting, but the numbers tell a competitive story that matters for anyone choosing between AI platforms.
Popular Android Emulator Allegedly Runs Hidden Surveillance Commands
A report circulating online claims that MuMu Player, an Android emulator from Chinese gaming company NetEase, allegedly runs 17 reconnaissance commands every 30 minutes without user knowledge. The specific evidence behind this claim isn't detailed in available materials. Community reaction on tech forums has been divided—some see it as confirming suspicions about Chinese software privacy practices and recommend running such tools only in sandboxed environments, while others invoke Hanlon's razor, suggesting poor engineering rather than surveillance intent.
Why it matters: For professionals using Android emulators for app testing or mobile workflows, this is a reminder to audit what background processes any emulator runs—and to consider isolation measures for software with unclear data practices.
Experienced Users Let AI Agents Work Twice as Long Without Interruption
Anthropic studied millions of human-agent interactions across Claude Code and its API to measure how much autonomy users actually grant AI agents. The findings: experienced users let agents run significantly longer without interruption—autonomous working time nearly doubled over three months, from under 25 minutes to over 45 minutes. Auto-approve settings jumped from about 20% among new users to over 40% with experience. Software engineering dominates agentic use at nearly 50% of API activity, with healthcare, finance, and cybersecurity emerging. Notably, on complex tasks, the AI stops to ask for clarification more than twice as often as humans interrupt it.
Why it matters: As users grow comfortable with AI agents, they're handing over more autonomy faster than oversight frameworks have developed—Anthropic's data suggests the industry needs new monitoring infrastructure before high-stakes domains like healthcare and finance scale up.
What's Innovative
Clever new use cases for AI
Mac App Helps Developers Run Multiple AI Coding Agents in Parallel
A developer released cmux, a macOS terminal app built specifically for managing multiple AI coding agent sessions—like Claude Code or Codex—running in parallel. The app uses vertical tabs that display git branch, working directory, and active ports, plus visual notifications when an agent needs your input. It's built on Ghostty's terminal rendering engine. Early testers on Hacker News called it 'pretty slick,' though some noted they'd stick with their existing terminal setup.
Why it matters: As AI coding assistants become workflow staples, expect more purpose-built tools to emerge for managing them—this is an early example of that trend.
Specialized Open-Source Model Targets Rust Programming
Fortytwo-Network released Strand-Rust-Coder-14B-v1, a 14-billion parameter model fine-tuned for Rust programming tasks. Built on the Qwen2 architecture, it's designed for code generation and conversational coding assistance in Rust specifically. No benchmarks or performance claims accompany the release.
Why it matters: This is developer tooling—relevant only if your engineering team works heavily in Rust and wants to experiment with specialized coding assistants beyond general-purpose models.
NVIDIA Releases Japanese-Language Model for Developers
NVIDIA released a Japanese-language version of its Nemotron Nano 9B model on Hugging Face. The 9-billion-parameter model is designed for text generation tasks and runs through standard transformer libraries. This is developer infrastructure—a smaller, specialized model for teams building Japanese-language AI applications.
Why it matters: For most readers, this is background noise—but if your organization serves Japanese markets and builds custom AI tools, this adds another option to evaluate against models from Japanese labs and larger multilingual systems.
Another Open-Source Language Model Joins Crowded Field
inclusionAI released Ling-2.5-1T, a text-generation model on Hugging Face. The model uses what the developers call a "hybrid architecture" for conversational AI, though no benchmarks or performance comparisons were provided. This is developer-oriented infrastructure—a new open-weights model joining the crowded field of downloadable language models available for custom deployments.
Why it matters: Without benchmark data or demonstrated capabilities, this is one of many model releases; worth noting only if you're actively evaluating open-source alternatives to commercial APIs.
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
OpenAI Grants $7.5 Million to Independent AI Safety Nonprofit
OpenAI committed $7.5 million to The Alignment Project, a nonprofit focused on independent AI alignment research. The funding aims to support work on AGI safety and security risks outside of major AI labs. OpenAI framed the grant as strengthening global efforts on alignment—the challenge of ensuring advanced AI systems behave as intended. No details were provided on specific research areas or how independence from lab interests would be maintained.
Why it matters: The grant signals that major AI labs are beginning to fund external safety research, though critics have long argued such work needs to be genuinely independent of the companies whose products it evaluates.
What's in Academe
New papers on AI and its effects from researchers
Top AI Models Score Below 10% of Human Performance on New Game-Based Intelligence Test
Researchers are proposing a new way to measure AI general intelligence: see how well models play human games. Their AI GameStore platform uses LLMs to generate test games based on popular titles from the Apple App Store and Steam. In a proof-of-concept with 100 games, seven leading vision-language models—including frontier systems—scored below 10% of human average on most games. The models struggled particularly with learning how game worlds work, remembering what happened, and planning ahead.
Why it matters: Current AI benchmarks often test narrow skills; this research suggests games could expose gaps in reasoning, memory, and adaptability that matter for real-world AI applications—and right now, even top models fail badly.
Math-Based Technique Steers AI Responses Without Full Retraining
Researchers introduced ODESteer, a framework that uses control theory mathematics to steer AI model behavior during text generation. The technique treats alignment—getting models to be more truthful, helpful, or less toxic—as a guidance problem solvable with differential equations. On benchmarks, it improved truthfulness scores by 5.7% over existing steering methods, with smaller gains on helpfulness (2.5%) and toxicity reduction (2.4%). This is research-stage work exploring how to fine-tune model outputs without full retraining.
Why it matters: If the approach scales, it could give AI providers a more principled way to adjust model behavior post-deployment—useful for compliance requirements or reducing harmful outputs without expensive retraining cycles.
Training Method Claims to Fix AI Reasoning Inefficiencies
Researchers introduced MASPO, a training method for improving AI reasoning that claims to fix inefficiencies in current approaches like GRPO (a technique used to train models on reasoning tasks). The paper identifies three technical problems with existing methods—wasted training signal, inflexible probability handling, and unreliable feedback—and proposes fixes for each. The authors claim "significant" improvements over baselines but provide no specific benchmark numbers in the available abstract.
Why it matters: This is deep ML research plumbing—if it holds up under scrutiny, it could eventually make reasoning models cheaper and faster to train, but there's nothing actionable here for non-researchers yet.
AI Web Agents That Predict When You'll Step In Score 27% Higher on Usefulness
Researchers developed an approach to make AI web agents more collaborative by training them to anticipate when users will step in to correct or guide them. They collected 400 real user sessions with over 4,200 mixed human-and-agent actions, identifying four distinct patterns of human intervention. Models trained on this data showed 61-63% better accuracy at predicting when users would intervene. In live testing, intervention-aware agents received 26.5% higher usefulness ratings from users.
Why it matters: As AI agents handle more complex web tasks—booking travel, filling forms, research—knowing when to pause and defer to the human could be the difference between a useful assistant and a frustrating one.
Open-Source Framework Aims to Simplify Scaling Recommendation Systems
Researchers released WarpRec, an open-source framework for building recommendation systems that aims to smooth the notoriously painful transition from research prototype to production deployment. The framework includes 50+ algorithms and can run the same code locally or distributed across servers without rewriting—a common pain point when scaling recommendation engines. It also tracks energy consumption during training. No performance benchmarks were provided comparing it to existing tools.
Why it matters: This is developer infrastructure—relevant if your engineering team builds recommendation systems in-house rather than using vendor solutions, but unlikely to affect most business users directly.
What's Happening on Capitol Hill
Upcoming AI-related committee hearings
| Tuesday, February 24 |
Building an AI-Ready America: Teaching in the AI Age House · House Education and the Workforce Subcommittee on Early Childhood, Elementary, and Secondary Education (Hearing) 2175, Rayburn House Office Building |
What's On The Pod
Some new podcast episodes
AI in Business — Improving Warehouse Efficiency with Unified Data and AI-Driven Visibility - with Dan Keto of Easy Metrics
The Cognitive Revolution — Mathematical Superintelligence: Harmonic's Vlad Tenev & Tudor Achim on IMO Gold & Theories of Everything