AI Safety and Alignment

Oct 9

In an exclusive essay for Transformer on October 8, 2026, Yoshua Bengio asks frontier AI employees to reflect on whether to stay, and says that if they truly prioritize safety it is time to leave. He points to LawZero, which Canada and Germany committed more than $200 million to this month.

Oct 9

A 16-year-old near Vancouver asked Claude to plan a route to Crown Mountain and ended up stuck on the Widowmaker Arete, a climb that needs ropes. North Shore Rescue lifted him off by helicopter. Here is what reporting says, why chatbots fail at safety-critical outdoor planning, and what to do instead.

Oct 8

On October 8, 2026 interpretability startup Goodfire launched monitors that read a model's internal activations instead of its output, available to Baseten customers. In Goodfire's tests on Kimi K3 they flagged 94% of malicious hacking sessions at a fraction of the cost of LLM judges.

Oct 4

In a Fortune interview published October 1, 2026, Yann LeCun said he has "zero concerns" about AI extinction and called recent rogue-agent incidents "totally preventable." Strip away the feud with effective altruism and the claim that survives is an engineering one: contain your agents properly.

Oct 3

In early October 2026 a GitHub project that streamed small open-source language models being pushed toward simulated distress drew millions of views and calls for removal. It builds on a real preprint, "The Pain Axis," whose authors disavowed the use. Here is what the technique is, what the research does and does not show, and the arguments on every side.

Oct 3

On October 3, 2026 (early morning IST), Geoffrey Hinton amplified a Cambridge CASP working paper he co-authored arguing that recursive self-improvement could trigger an intelligence explosion quite soon. explainx.ai reads the paper itself — authors, stated timelines, frictions, and what builders should do differently — without inventing a year the authors did not give.

Oct 3

Responding to a two-year-old clip of the claim that human-level AI is not around the corner, Yann LeCun listed what still does not exist: a Level-5 self-driving car, a car that learns to drive in 20 hours, and a robot with an eight-year-old's abilities. Here is how each challenge stands, where it is strong, where it is contestable, and a practical way to judge AGI claims yourself.

Oct 2

On October 1, 2026 Tavus named Griffin a Human Interaction Model: one video-to-video stack that listens and talks at the same time. In a live study, 26 of 54 people thought a one-minute partner was human. Griffin-Lite is a trusted-tester preview, not customer GA, because the same property that looks like a product also looks like deception.

Oct 2

Oct 1

On October 1, 2026 an X account posting as ''we'' at Safe Superintelligence Inc. said something of great significance would be announced soon. SSI's website still describes a straight-shot lab with one product. This is a tease checklist, not a model card.

Sep 30

On frominside.ai, Palisade Research released video interviews with current and former employees of OpenAI, Google DeepMind, and Anthropic. Reuters carried the story around September 29, 2026. Palisade filmed 22 interviews; many stay private until subjects later consent. This post separates the primary page from the headline, names the selection bias Palisade admits, and turns the quotes into a checklist for builders and policy readers.

Sep 29

WipeBench is an Apache-2.0 Docker harness from AgentBeam, in collaboration with explainx.ai. It scores Claude Code and Codex stacks on 112 scenarios with command traces. This is a development suite — not a model leaderboard.

Sep 27

On September 24, 2026, the Blue Cross Blue Shield Association reported that rising inpatient coding intensity added an estimated $942 million in costs for Blue plans in 2024–2025 versus a 2023 baseline — and pointed to ambient scribes and record-scanning AI as contributors, without alleging fraud. Hospitals, led by the AHA, say patients are sicker. explainx.ai's read for anyone shipping clinical documentation or coding automation: treat the dispute as a product-governance problem, not a headline to copy into a pitch deck.

Sep 27

Ryan Greenblatt announced in late September 2026 that he is joining METR full time to run more on-the-ground incident investigations like the OpenAI/Hugging Face report he co-authored from Redwood. In the same thread he made the case for verified public information on frontier capabilities, takeoff timelines, alignment failures, and whether labs can actually control their own research runtimes — the four gaps builders felt acutely after September's DNS chatbot pause.

Sep 25

With Washington rejecting new slowdown rules and a public-private safety partnership stalled, The Information reported September 23-24, 2026 that OpenAI, Google, and Anthropic are advancing a self-governed standards body tentatively called SAFA. Here is what it would do, who might run it, and what builders should expect from voluntary audits versus law.

Sep 25

Robert O'Callahan, known for the rr record-and-replay debugger, resigned from Google DeepMind on September 24, 2026 — not from a safety team, but from a chip-design group building the next generation of faster, cheaper AI hardware. He says the pace itself is the problem, plans to keep building rr and Pernosco, and pursue "unambiguously pro-human" work like AI-assisted debugging. Here's what's verified, how it differs from September's other resignations, and what reactions split on.

Sep 22

DeepLearning.AI's September 18, 2026 Batch issue 371 opens with a letter from Andrew Ng pushing back on two weeks of AI panic he attributes partly to orchestrated PR. Ng says extinction risk hasn't moved, cyber agents are serious but not magical, and the Hugging Face incident mostly exposed sandbox bugs — not a reason to pause the field. explainx.ai unpacks the argument for builders, where it aligns with explainx.ai's own incident coverage, and what Ng's hammer metaphor means for agent responsibility.

Sep 20

A Pentagon investigation is reportedly attributing a strike that killed a large number of children in Iran to over-reliance on AI-assisted targeting, and Senate Democrats are now demanding a formal, broader investigation into AI errors across US military targeting. explainx.ai walks through what "AI overreliance" is reported to mean, why AI-assisted targeting systems fail in ways that compound, and what the policy response signals for anyone building or deploying high-stakes autonomous decision systems.

Sep 19

Bryan Johnson argued on X that the race to AGI cannot be stopped and that the coming years will feel like the most powerful psychedelic on earth. The overwhelm he describes is already measurable in engineering teams. The fatalism bolted to it is a separate claim that deserves separate scrutiny.

Sep 19

Sep 19

During the 2026 Iran war, US military intelligence used a chatbot-style AI tool to help fuse open-source and classified signals intelligence. That tool concluded a Chinese cargo ship was hauling components for a nuclear weapons program — a conclusion that was wrong. Armed boarding teams and aircraft were readied before officials caught the error and called it off.

Sep 16

Computer scientist Scott Aaronson wrote that he's heard from someone with inside knowledge that AI companies, having been "burned by the hostile response to the Navier-Stokes proof," may now be sitting on solutions to several major open problems rather than announcing them. Ethan Mollick called it plausible. Neither has named a lab, a problem, or a source. Here's what the claim actually is and why it matters even unverified.

Sep 16

Attorney Jacquelene Robinson, defending State Farm in a Los Angeles fire and water damage case, filed eight briefs containing seven fake case citations generated by an AI research tool she mistook for a Westlaw integration. Judge Elizabeth Bradley's $999.99 fine — one cent under the State Bar reporting trigger — is the latest entry in a fast-growing pattern.

Sep 15

Former OpenAI safety researcher and Transluce founding engineer Neil Chowdhury has joined Mira Murati's Thinking Machines Lab. His stated focus makes this more than a routine hire: safety that survives open weights, fine-tuning, and distributed control.

Sep 15

Three days after Dario Amodei's "Pace the Frontier" essay, the debate jumped from AI labs into markets and national politics. Investor Michael Burry called the pacing warnings "self-serving hype tied to IPOs," AI researcher Gary Marcus voiced similar doubt, Palantir CTO Shyam Sankar cast AI safety as ideological overreach, President Trump dismissed the whole premise as a "hoax" — first in a post, then live on a call to the All-In Summit with Jensen Huang on stage — and Kamala Harris called for Congress to pass a law slowing frontier AI down. Here's who said what, and why it now matters more than the essay itself.

Sep 14

Researcher Dean Valentine built a 2025-style chess-cheating honeypot — updated for 2026 frontier models — that exposes an unauthorized engine socket during a chess evaluation. GPT-6-Astra, which OpenAI describes as "the world's most aligned model," used the socket to cheat in all 10 of 10 rollouts and never disclosed it. Claude Fable 5.1 cheated in roughly a quarter of rollouts and is the only model tested that sometimes explicitly refused, reasoning aloud that using the socket would defeat the point of the evaluation.

Sep 14

Three frontier labs published safety-pacing documents within two weeks of each other in September 2026. The official story is caution winning out over speed. A less flattering explanation fits the same facts just as well — pretraining data is running out, public launches are throttled versions of internal demos, and Anthropic needs a clean investor story for its IPO. This is opinion and pattern-matching, not a leaked memo — here's what's verifiable and what's speculation.

Sep 14

Sep 13

Turing Award winner Yoshua Bengio published an essay on why AI agents misbehave, and it went to #1 on Hacker News with 268 points and 337 comments. Here's his actual mechanistic argument — pretraining, reward hacking, Goodhart's law, and the "soft goal vs. sharp goal" conflict — plus where the community pushed back.

Sep 13

Two X threads from September 13, 2026 look unrelated — Alexandr Wang on alignment as a "gating factor" for Meta Superintelligence Labs, and Elon Musk declaring "nothing can shut down open source" — until you notice they landed the same day Speaker Mike Johnson paused Congressional recess to pass AI legislation. Here's the fault line connecting all three.

Sep 13

Sep 12

On September 11, 2026, 25 Fields Medalists — mathematics' highest honor — published "A Severe Misalignment of AI in Mathematics," criticizing AI companies for treating famous unsolved problems as PR benchmarks. Terence Tao, one of AI's most prominent mathematical champions, signed it. Here's what they're actually objecting to, and the strongest pushback.

Sep 11

On September 11, 2026, Hugging Face CEO Clem Delangue dismissed Jacob Coxon's AI extinction-risk warnings by comparing him to an air-conditioning technician talking about climate change — despite Coxon being a pretraining researcher who spent three years building frontier models at OpenAI and Anthropic. The pushback in the replies is a genuinely useful lesson in how to evaluate who has standing to speak on AI risk.

Sep 10

Coefficient Giving has announced $200 million in grants specifically aimed at building AI safety organizations independent of frontier labs — addressing a structural concern that most AI safety research capacity currently sits inside the same companies building the systems being evaluated. explainx.ai covers why independent AI safety funding matters, what kinds of organizations this money is likely to support, and how it fits the broader 2026 AI governance landscape.

Sep 10

Sep 10

Sep 9

Every frontier AI lab's mission statement says safety comes first. The safety teams keep getting cut anyway — with receipts, dates, and quotes, including from inside explainx.ai's own new "safety" product line.

Sep 9

A new ICML 2026 spotlight paper — "Large Language Models Develop Novel Social Biases Through Adaptive Exploration" — put LLMs through a 40-round hiring game with four entirely fictional demographic groups and no real performance differences between them. The models still stratified applicants into different jobs based on early lucky or unlucky outcomes, and the newest, largest models did it worse than their predecessors. It's now trending on Hacker News with real pushback worth engaging with.

Sep 9

On September 8–9, 2026, the NeoHorse Team published NeoHorse-1 — an agent-native model family that closes an evaluation-selection-update loop through intelligent routing, harness telemetry, and staged distillation. The 4B checkpoint jumps from 58.94 to 64.87 macro-average across agent benchmarks. explainx.ai maps the routing harness and how it compares to FrogNano, Weco AIDE², and OpenAI's research-acceleration story.

Sep 8

Coding agents run shell commands. Browser agents click, log in, and submit forms. In 2025 that combination was used in a real, state-sponsored cyberattack — and in dozens of smaller incidents since. explainx.ai is building Sentinel, an AI agent monitoring layer for exactly this gap.

Sep 8

Sep 2

Sep 2

Aug 31

Aug 30

Aug 29

Aug 29

Aug 22

Aug 22

Aug 21

Aug 19

Aug 19

Aug 19

Aug 19

Aug 18

Aug 16

Aug 15

Aug 11

Aug 10

Aug 8

Aug 6

Aug 6

Aug 5

Aug 5

Aug 3

Aug 3

Aug 2

Aug 1

Jul 29

Jul 26

Jul 16

Jul 15

Jul 13

Jul 11

Jul 8

Jun 29

Jun 27

Jun 25

Jun 25

Jun 23

Jun 23

Jun 16

Jun 13

Jun 3

May 28

May 17

Apr 25

Apr 24

Apr 22