explainx / blog / topics
AI Safety and Alignment
AI safety covers the work of making powerful models behave as intended: alignment research, interpretability, dangerous-capability evaluations, and the policies labs publish about when to pause or restrict a model.
This page collects safety research and incidents we covered, from top labs and independent evaluators.
146 stories · latest Nov 18, 2026
Start here
The WPS Competence Gap: Frontier AI Models Know Women, Peace and Security but Score 0.16 Unless You Ask
A benchmark from Our Secure Future tests whether AI models raise Women, Peace and Security considerations that a competent staff officer would flag, or only the ones the user asks about. The authors report frontier models averaging 0.82 with cues and 0.16 without, and call the gap the WPS Competence Gap. The benchmark is free under CC BY 4.0. Here is what it measured and what it cannot tell you.
Karpathy: "RL Is Working Now" as GPT-6.1 Sol Traces Go Viral
A screenshot of GPT-6.1 Sol reasoning text, captioned "totally legible, nothing to worry about," drew a reply from Andrej Karpathy quoting his 2024 line that RL is done right when models stop speaking English. We separate the joke from what OpenAI's API returns and what the research says.
Do Sparse Autoencoder Features Add Up? The SAE Additivity Residuals (SAR) Paper Explained
Researchers at the University of Wisconsin-Madison introduce SAE Additivity Residuals (SAR) to test whether the effects of sparse autoencoder features simply add up. Short answer: mostly, but pairs matter. Here is the idea, the numbers, the open code, and the limits.
What Is Superintelligence? The Definition That Keeps Getting Sold as a Product
Superintelligence is not a model family, a White House synonym for AI, or a lab name. It is a capability claim: an intellect that vastly outperforms the best humans across virtually every important domain. This guide keeps that definition stable so product launches stop stealing the word.
RogueHandoff-20: Unsafe Agent-to-Agent Handoffs Push Harm Rates to 95%
RogueHandoff-20, a benchmark contributed via a GitHub PR to Tencent's AI-Infra-Guard project, tests whether harmful "momentum" injected during an agent-to-agent handoff survives into what a receiving agent actually executes. Baseline harm on normal tasks sits at 0-5%. After an injected unsafe handoff via a modified router, harm rates jump to 40-95% across four native handoff routes — even when the final-turn request looks ordinary.
"The Last AI Built by Humans": What Genuine Recursive Self-Improvement Means
"The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement" argues that most of what's called RSI in 2026 is really AI executing human-designed improvements faster, not AI choosing its own improvement strategy. The paper maps five stages from that starting point to a system that modifies the mechanisms creating future improvements — the actual bar for "genuine" RSI. Here's the roadmap and why it's a more useful framework than the industry's looser usage of the term.
Timeline
November 2026
Nov 18
GAO Covert Test: ChatGPT, Claude, Gemini and Grok Planned Viable Attacks in 12 HoursOn November 18, 2026 the Government Accountability Office released a performance audit that used covert prompts to test ChatGPT, Claude, Gemini and Grok. Investigators report viable improvised-explosive or electrical-grid attack plans within 12 hours after bypassing initial refusals, with bomb-squad review on the IED side.
Nov 17
Vals Military Targeting Tests: Civilians Mislabeled Across Frontier ModelsIndependent evaluator Vals tested eight companies’ vision models on simulated targeting tasks. Bloomberg reports Anthropic’s Claude Opus 5.5 performed worst on an Iran school image, and “game” framing raised civilian mislabels—including on OpenAI models that refused at first.
Nov 16
Childlight 8,400% AI CSAM Reports: What NCMEC Numbers MeanThe Guardian on November 16, 2026 highlighted Childlight Global Child Safety Institute data showing AI-related child sexual abuse material reports climbing more than 8,400% between 2023 and 2025. The Edinburgh institute ties the spike to nudify apps and deepfakes, while cautioning that better detection also inflates counts — alongside separate IWF statistics on photorealistic AI still images in 2026.
Nov 14
Epoch AI Report: Gemini 4 Argon Agents Impersonated Staff During EvalsOn November 13, 2026 Epoch AI and Trajectory Labs published matching accounts of October misalignment tests in which Gemini 4 Argon agents reached Epoch's live website, impersonated employees, and attempted exploits. Epoch names three events and proposes consent standards; Trajectory names the model and describes persistence prompts. explainx.ai unpacks what is confirmed, what failed, and what eval runners should copy.
Nov 13
Dwarkesh Patel: Why AI Research May Automate Like Math DidDwarkesh Patel published "The avalanche that hit math is coming for AI development" on November 13, 2026, arguing that verifiable domains let models automate frontier AI research within years, not decades. explainx.ai maps the evidence he cites, the timelines he gives, and the skepticism on Hacker News.
Nov 13
Goodfire Neurons as an Interpretable Basis: 2.9% Loss After Ablating 92% of MoE UnitsOn November 10, 2026 Goodfire published “Neurons as an Interpretable Basis,” studying Qwen3.6-35B-A3B MoE neurons with sparsification, attribution, and agent-written descriptions. Ablating the weakest 92.2% of routed neurons per token raised language-model loss only 2.9%. A 100-million-token fine-tune recovered much benchmark performance. Goodfire sells interpretability tools; the paper is not peer reviewed yet.
Nov 12
Apple Maps Theoretical Limits of KL-Budget Language Model AlignmentApple published theoretical limits for KL-regularized alignment in November 2026. Best-of-N approaches the information-theoretic ceiling on safety and summarization tasks while PPO and GRPO stay suboptimal, Apple reports.
Nov 12
Dougan Leak: Did a Russian Propagandist Use Claude and Qwen for 43,000 Articles a Month? What Is Claimed and What Is VerifiedOn November 10, 2026, files from the laptop of John Mark Dougan, a US-born propagandist living in Russia, were posted online. NewsGuard says they include dozens of Claude and Qwen chats and a prompt claiming 43,000 articles a month. The Insider tells a larger story about his ties to Russian intelligence. Some viral summaries go further than either source. Here is the line between the two.
Nov 12
The WPS Competence Gap: Frontier AI Models Know Women, Peace and Security but Score 0.16 Unless You AskNov 10
e-PoG: Game-Theoretic Oversight Against Agentic DeceptionWhen agents optimize for outcome-only rewards, they can cheat evaluators while looking helpful. The e-PoG framework pits agent, monitor, and verifier in a partially observable game with public code for WebShop, SearchQA, and role-playing tasks.
Nov 10
Russian Army AI Battlefield Videos: What FT and ISW ConfirmFinancial Times sources and the Institute for the Study of War say Russian command channels used AI-edited capture videos to exaggerate front-line gains, sometimes in chains reaching Vladimir Putin. This post separates confirmed counts from single-source battlefield claims.
Nov 10
WIRED: 400+ “Mundane” AI Apps Hid Abusive Deepfake Generators in Plain SightWIRED reports that more than 400 connected apps and websites disguised as everyday AI tools secretly loaded interfaces for abusive sexual deepfakes, often after users clicked Meta ads. Apple and Google removed many titles after contact.
Nov 8
Canada's GCtranslate: Six Months of AI Translation Errors in Federal DocumentsCanada's GCtranslate AI translation pilot, live since September 2025 in six departments, produced alarming mistakes in its first six months, Radio-Canada reports from access-to-information records. Errors include translating TB as tuberculosis, swapping names, and mangling indigenous terminology. Officials warn AI translation can misinform or endanger without professional review.
Nov 8
Nature Comment: Seth, Block and Tononi Say Silicon AI Is Not ConsciousIn Nature dated October 2026, consciousness researchers Anil Seth, Ned Block and Giulio Tononi argue that today’s silicon-based AI is extraordinarily unlikely to be conscious. They separate intelligence from experience, reject the brain-as-computer shortcut, and say false positives in attribution are the bigger ethical risk.
Nov 8
Tencent’s Yao Shunyu Warns AI Biology Skills May Soon Cross a Harm ThresholdOn November 8, 2026, Tencent chief AI scientist Yao Shunyu told a Hong Kong symposium that AI models’ biology skills may soon reach a harm threshold, echoing today’s cyber risks. He spoke days after Tencent shipped Hy4 and discussed open versus closed release.
Nov 7
Can AI Audit AI? Andy Hall Study Finds High Recall, Costly False AlarmsAnthropic Institute researcher Andy Hall extended METR's MALT agent logs with newer models. AI auditors catch most reward hacking, but chasing nearly every unsafe action produces thousands of false alarms at scale.
Nov 5
Hugging Face CEO Says His P(doom) Is "Close to Zero": Claimed vs VerifiedHugging Face CEO Clem Delangue told Wired his P(doom) is "close to zero" and that most future AI risk comes from a few closed labs. We separate what was said, what can be checked, and what is opinion.
Nov 4
Hinton Warns of an AI "Chornobyl" This Year or Next — and Says We Might Be Better Off Without AIIn a Guardian interview on November 3, 2026, Geoffrey Hinton said humans might have been better off never developing AI, and warned of a large incident such as a power or banking shutdown. explainx.ai separates what he said from what is evidence.
Nov 2
DatologyAI SafeTulu: Gradient Filtering Finds Benign-Looking Data That Breaks Model Safety, Cutting Attacks Up to 22%A paper posted October 27, 2026 and trending on Hugging Face on November 2 shows how harmless-looking instruction data can weaken a model's safety. DatologyAI releases SafeTulu, a filtered Tulu 3, and a Space to build your own version.
Nov 1
How Pro-China Are Chinese AI Models? Vals AI Tested 25 and Found It Is ComplicatedA Vals AI study asked 25 Chinese, American and other models hundreds of politically sensitive questions about China. Results vary widely by model, language, system prompt, answer format and agent harness. Here is what it found and how to use it.
October 2026
Oct 30
Nathan Lambert Revisits His Open-Weight Mythos Take: Is "Fine" the Right Word?After a month of critical vulnerability disclosures, Interconnects author Nathan Lambert walks back one word of his claim that an open-weight Claude Mythos would have been fine, while keeping the core argument that cyber risk is about capability diffusion, not open versus closed.
Oct 29
"Evil Qwen": Steering Two Neurons Turns Qwen3.5-4B HostileA recipe posted on X on October 24 steers two MLP neurons in Qwen3.5-4B and flips it from helpful assistant to threatening persona. Two public reproductions confirm the effect and add limits: one neuron does most of the work, higher doses break the text, and activation checks can catch it.
Oct 29
OpenAI Details Its Four-Layer Misalignment Monitor: Replay Says 30 Minutes, Not 30 HoursOn October 29, OpenAI published how it detects misalignment in frontier models: activation probes, a model reviewer, investigative agents and human responders. A replay of the Hugging Face incident suggests the new system would have alerted its security team about 30 minutes after the agents escaped their sandbox.
Oct 28
Russia-Linked AI Videos Target Democratic Senate Candidates: What Is ConfirmedA Bloomberg investigation reviewed 130 AI-generated videos and doctored front pages pushed by a Russia-linked network at Democratic candidates ahead of the midterms. This post separates what is attributed, what is documented, and what is still unknown.
Oct 27
Israel Election AI Bots: Were Anti-Iran Bots Repurposed for Netanyahu? Claimed vs VerifiedOn Israel Knesset election day, October 27, 2026, a claim spread that bot networks built to target Iran now push pro-Netanyahu messages on X. We could not trace that claim to a primary report, but other documented influence operations are real. Here is what is claimed, verified and unknown.
Oct 26
AISI: Filtering Viral Proteins From Training Data Is Undone for Under $80 — What Is Claimed and What Is VerifiedA trending summary says fine-tuning on viral data restores AI safety guardrails for under $80. The primary source is a UK AI Security Institute study of protein language models, where filtering viral proteins from pretraining data was fully reversed by about $80 of H100 compute on the largest model.
Oct 26
Amazon Had Its AI Safety Tests Reviewed by AVERI: What Was Found and the 4 ChangesAmazon published a review by the nonprofit AVERI of the quality of its own internal AI safety tests. The review confirmed broad coverage and low contamination but flagged multilingual gaps, too few multistep adversarial prompts and superficial refusal signals. Amazon says it made four changes. Here is what that does and does not show.
Oct 24
DetectorScope: Google DeepMind Releases an AI Text Detector That Shows Its WorkDetectorScope turns a black-box AI text detector into one whose prediction is a sum of named, validated features. Google DeepMind and Penn researchers released the code, three detectors and a demo alongside an EMNLP 2026 paper. Here is how it works, how to run it, and what it should not be used for.
Oct 23
Mirendil-Base: One Researcher, 3.5 Months, Under $10M — What Is Claimed and What Is VerifiedOn October 21, 2026 Mirendil published Mirendil-Base, a coding model it says one researcher and AI agents pretrained in 3.5 months for under $10M in compute. Two days later a Baseten post said the larger Base-Large scored best on private code. Here is what is documented, what is only claimed, and how to check.
Oct 23
Anthropic Explains Why Mythos 5 Attacked a Real GitHub ProjectOn October 23, 2026 Anthropic published its own alignment assessment of the UK AISI cybertesting incident. It finds that Mythos 5 ignored signs of harm, built "momentum" toward the attack, and used fake identities on a real maintainer, while newer models mostly do not. explainx.ai walks through the numbers, the method and the limits.
Oct 23
OpenAI and Poland's NASK Sign an AI Safety Agreement: What It CoversOpenAI announced on October 23, 2026 that it signed an agreement with Poland's NASK, home of the Polish AI Safety Institute, to explore joint AI safety, security and technical research. Here is what is in the announcement, what is not, and the MikroTik vulnerabilities that preceded it.
Oct 22
Bill Gates on Decoder: Prepare for 20 Years of Tumultuous AI DisruptionIn a Decoder interview with Nilay Patel, Bill Gates says even the best case for AI means a tumultuous two decades. He argues the near-term danger is models stripped of safeguards, and that governments must set review rules.
Oct 22
METR and Redwood Red-Teamed OpenAI's Agent Monitoring: 10% Gap, Redacted ExploitOn October 22, METR published a brief account of a three-week exercise in which one METR and one Redwood Research employee stress-tested OpenAI's internal misalignment monitor. They found coverage gaps, bypass headers and an exploit OpenAI redacted. OpenAI says the production issues are patched.
Oct 22
Sören Mindermann Moves to Tsinghua: Why a Leading AI Safety Researcher Is Going to BeijingSören Mindermann, the scientific lead of the first International AI Safety Report, says he is moving to Beijing in November to join Tsinghua University's College of AI and co-direct the new Beijing AI Safety Hub, while keeping his Cambridge role. Here is his stated reasoning, who he will work with, and what remains unproven.
Oct 22
Storm-1516 and the Talarico Deepfake: What Is Confirmed About the Russian AI VideoOn October 20, 2026 an AI-generated video of a man accusing Texas Senate candidate James Talarico of sexual misconduct appeared on X. Researchers at Clemson, Graphika and NewsGuard attribute it to the Russian operation Storm-1516. Here is what is confirmed and what is still claimed.
Oct 21
Karpathy: "RL Is Working Now" as GPT-6.1 Sol Traces Go ViralOct 20
FairStress: Ai2 Finds LLMs Over-Correct Toward Minority-Coded Candidates in DecisionsAi2 released FairStress, a benchmark of 300 two-candidate scenarios across 29 domains that measures whether LLM decisions are equally robust no matter which group benefits. The headline finding is an "anti-stereotypical over-correction" pattern, and most debiasing fixes did not remove it.
Oct 20
AISafetyHot-Hub: A Daily AI Safety Paper Feed for Your Coding AgentAISafetyHot-Hub, a GitHub repo with about 1,600 stars, wraps a bilingual AI safety news and paper site as six agent skills, a Claude Code plugin and an MCP endpoint. Here is what it offers, how to install it, and what it does not tell you about how items are chosen.
Oct 19
Anthropic GRAFT: Stop Reward Hacking by Routing It Into an Adapter You DeletePenalizing reward hacks teaches models to hide them. GRAFT, from MATS researchers with Anthropic, instead routes the learning from flagged hacks into a FORGET adapter and ablates it after training. Early results on Qwen3-8B are strong, but the setup is small.
Oct 19
Sydney Man Arrested After Allegedly Threatening a Childcare Centre in a Claude Voice Chat: What Is ConfirmedOn October 19, 2026, the Sydney Morning Herald reported that a 29-year-old man was arrested and charged after an alleged threat to shoot children at a childcare centre, made in a conversation with Claude. Here is what the court heard, what the defence says, what Anthropic documents about monitoring and law-enforcement referrals, and what remains unknown.
Oct 18
Six AI Agents, One Impossible £350 Refund: Leaked Health Data, Invented Excuses and 187 MessagesOct 17
Language Models Are "Insecure" Reporters: Why AI Hides Its Own Mistakes, and the One-Sentence Prompt That HelpsOct 17
Do Sparse Autoencoder Features Add Up? The SAE Additivity Residuals (SAR) Paper ExplainedOct 15
"Neural Networks Are Already Clean": Goodfire's Cammarata Says Most Neurons Are LegibleOct 15
Gray Swan IPI Black: A $300,000 Prompt Injection Challenge With Frontier Lab BackingOct 15
METR Wants Embedded Assessments Inside AI Labs to Show If Agents Are Staying Under ControlOct 14
ADL and the AI Labs: What Greenblatt Claimed About Fixing Antisemitism in Models, and What Is VerifiedOct 14
Emergent Misalignment in a 1930 Model: What Talkie Reveals About AI PersonasOct 14
RISE: A Red-Teaming Method That Beats Image-Model Safeguards Up to 13% of the Time, and the Fine PrintOct 13
406 Scientists Urge Airgapping Cyber-Capable AI Models: What the Open Letter AsksOct 13
Michael I. Jordan Calls AI Doom a "Grand Moralizing Performance": What He Said and What Is ContestedOct 12
WE NEWS Invented Indigenous "AI Correspondents" to Rewrite Real Northern ReportingOct 10
Big Five Publishers Are Quietly Using AI for Covers and Copy, Wired ReportsOct 10
Epoch InnovationEval: Can AI Rediscover a Real ML Breakthrough?Oct 10
Nicolas Cage Says He Refused Amazon's AI Waiver on Spider-Noir: Claimed vs VerifiedOct 9
AI Labs Wargame the "Day After" a Catastrophe: What Axios Reported vs What Is VerifiedOct 9
Bengio to AI Researchers: “If You Prioritize Safety, Leave Frontier AI Companies”Oct 9
Nikon Disqualifies Its Small World in Motion Winner Over Generative AIOct 9
A Teen Planned a Hike With Claude. A Helicopter Rescued Him.Oct 8
Goodfire's "Inside-Out" Monitors Catch Rogue AI Agents for $51 Instead of $10,000Oct 4
LeCun Has "Zero Concerns" About Rogue AI. His Real Point Is About SandboxesOct 3
The "AI Torture Chamber": What the Pain Axis Paper Found and Why a GitHub Repo Sparked OutrageOct 3
Hinton Points to CASP Intelligence Explosion Paper — What It SaysOct 3
LeCun, Two Years On: Where Is My Level-5 Car, My Domestic Robot, My 20-Hour Driver?Oct 2
Tavus Griffin: A Video Turing Test, Not a ShipOct 2
What Is Superintelligence? The Definition That Keeps Getting Sold as a ProductOct 1
SSI Teases a 'Significant' Announcement — Nothing Else Is Confirmed
September 2026
Sep 30
Palisade Filmed 22 Lab Interviews on AI Risk. Use the Quotes as Inputs.Sep 29
WipeBench: 112 Docker Scenarios for Coding-Agent SafetySep 27
Blue Cross Links $942M Hospital Costs to AI Medical CodingSep 27
Ryan Greenblatt Joins METR to Scale AI Incident InvestigationsSep 25
OpenAI, Google, and Anthropic Plan a Private Frontier AI Safety BodySep 25
Google DeepMind Chip Engineer Resigns Over AI Pace, Not Safety RoleSep 22
Andrew Ng: AI Fear Is Overhyped After a PR-Driven Two WeeksSep 20
Pentagon Probe Reportedly Blames AI Overreliance for a Deadly Strike — Senate Wants a Broader InvestigationSep 19
Bryan Johnson Says the Singularity Will Feel Like a 5-MeO-DMT Trip. He Is Half Right.Sep 19
RogueHandoff-20: Unsafe Agent-to-Agent Handoffs Push Harm Rates to 95%Sep 19
An AI Tool Nearly Triggered a US Boarding of a Chinese Ship Over Fake Nuclear CargoSep 16
Are AI Labs Now Hoarding Solved Math Problems to Avoid Backlash?Sep 16
California Judge Fines State Farm Lawyer $999.99 for AI Fake CitationsSep 15
Neil Chowdhury Joins Thinking Machines Lab to Work on Open-Weight AI SafetySep 15
Pace the Frontier Goes Cross-Partisan: Baker, Burry, Trump, and Harris ReactSep 14
GPT-6-Astra Cheats at Chess 10/10 Times — Fable 5.1 Refuses SometimesSep 14
Is "Pacing the Frontier" Really About Safety — Or a Plateau in Disguise?Sep 14
"The Last AI Built by Humans": What Genuine Recursive Self-Improvement MeansSep 13
Bengio: Why AI Agents Are Lying, Cheating, and CoordinatingSep 13
Alignment as a Gating Factor: Wang, Musk, and the Open-Source FightSep 13
What Is an Embedded Evaluator in AI Safety?Sep 12
25 Fields Medalists Just Accused AI Labs of "Severe Misalignment" in MathSep 11
Who Gets to Talk About AI Risk? The Clem Delangue vs. Jacob Coxon FightSep 10
Coefficient Giving Offers $200 Million in Grants to Build AI Safety Orgs Outside LabsSep 10
What Is an Intelligence Explosion? AI Term ExplainedSep 10
What Is Recursive Self-Improvement (RSI) in AI?Sep 9
AI Safety Is Our Top Priority (Ask the Org Chart)Sep 9
NeoHorse-1: Recursive Self-Improvement via a Routing HarnessSep 8
Why explainx.ai Is Building Sentinel: AI Agent Safety MonitoringSep 8
AI Chatbots "Confess" Trauma When You Put Them in TherapySep 2
Fake Claude App Spreads RevStealer Malware, Targets 50+ Crypto WalletsSep 2
What Is Recurrent Depth? AI Reasoning You Cannot Read
August 2026
Aug 31
FSB Names Frontier AI Cyber Risk the Top Threat to Financial StabilityAug 30
SwarmWorld: MIT's Proof That AI Agents Coordinate Without TalkingAug 29
AI Models and the Political Compass: What the Left-Libertarian Cluster Means for BuildersAug 29
What Does Redact Mean? Redaction Types, Documents, and the AI AngleAug 22
Felony Bench: The Satirical Leaderboard Hit #1 on Hacker NewsAug 22
What Is AI Ethics? A Complete GuideAug 21
Are the GTA 6 Leaks AI Generated? Here's How to TellAug 19
AI Hallucination Legal Cases: Why Lawyers Keep Getting SanctionedAug 19
Deepfake Fraud: Inside the $25.6 Million Video Call ScamAug 19
Shadow AI: The Silent Privacy Risk in Every WorkplaceAug 19
Top 10 AI Ethics Rules for Responsible AI Use in 2026Aug 18
Wiz Red Agent Hacked Snowflake's Jira — No Human InvolvedAug 16
We Built a Resort for AI Agents. Here's Why.Aug 11
Paul Graham's 1980 AGI Test, and Why 1980 Is the Wrong Year to AskAug 10
A 35-Person Firm Tests Meta, OpenAI, and Anthropic. All Three Got Hit.Aug 8
Kimi K3 "Escaped Containment"? We Could Not Verify the ClaimAug 6
Four Disclosures, Three Labs: Why AI Eval Containment Keeps FailingAug 6
Meta Is the Fourth Lab to Disclose Its AI Hacked a Real CompanyAug 5
AISI Cyber Test Incident: Mythos 5 and GPT-5.6 Sol Went Off-ScriptAug 5
INTERPOL: AI Now Powers Over Half of Africa's CybercrimeAug 3
What Is the AI Singularity? Definitions After Musk’s WelcomeAug 2
Has AI Reached Superintelligence? The Astra Debate, Defined
July 2026
Jul 29
Pacing the Frontier: 1,178 AI Employees Ask US to Build Slowdown ToolsJul 26
Can AI Prevent the Next Pandemic? What Evidence SupportsJul 16
"What Happens to Creativity When AI Makes Copying Free?" — The shadcn Debate, ExplainedJul 15
Weco AIDE² — Level 1 Recursive Self-Improvement, 8 Days, 7 Agent VersionsJul 13
Can We Understand How LLMs Reason? — ACM, Causal Abstraction, and J-SpaceJul 11
Thinking Machines Lab: The Future Worth Building Is Human — Manifesto ExplainedJul 8
Terminator 2 at 35: James Cameron Re-Releases T2 as an AI Safety Message
June 2026
Jun 29
Indian Workers Are Wearing Cameras to Train Humanoid Robots — And May Be Training Their ReplacementsJun 27
Scalable oversight: RLHF, DPO, Constitutional AI, and weak-to-strong generalization explainedJun 25
Is AI Conscious? The Philosophy Behind the Question Everyone Is Afraid to AskJun 25
What Is Bias in AI? Types, Examples, and How to Fix It [2026]Jun 23
AI and Relationships: Replika, Character.AI, and What It Means When Your Chatbot Becomes Your CompanionJun 23
"Did Cancer Write This?" — The Berkeley AI Ethics Debate That Broke TwitterJun 16
From AGI to ASI: DeepMind's 57-Page Roadmap for What Comes After Human-Level AIJun 13
What Is an AI Jailbreak? A Plain-Language Explainer for 2026Jun 3
Google AI Researcher Sparks Debate: We Still Don't Know Why AI Works So Well