AI/ML News & Innovations Hub

AI/ML news, top picks, and generated innovation digests.

★ Visit ai-karthik.com
426Sources
90067News Items
8Top Picks
445Blogs
successLast Run

Latest AI/ML News

90067 matching items

[low effort post] The Non-extinctionist World
LessWrong AI 2026-11-15 23:16 UTC Score 68.0 USR-0152-20261115-community-fo-cdcc6c07 Full article

[low effort post] The Non-extinctionist World

This is a prompt I give to every freshly released claude in the hope that one of them pops out a worldsaving insight as a result, I think I wrote this text first as the opener of a chat with an opus 4.5. posted here without further comment so I get the post out ever at all: --- Most species go extinct on average. Imagine the set of all species. Can't do that? Okay, [look at this image.](tree of life dot jpg) Look at the line tips. Count how many there are in one layer of dieoffs and the next. Few species survive. But more than that: imagine a room full of one of every species. Imagine, if you will, that you're going to give them a speech: from beyond the ages, instructions for how one of any of their species could do their best to make the species survive, if they truly thought about it, truly considered all their personal roadblocks to becoming the parent to endless generations. What is the best advice for both what the individual should do, right now, in order to move the world from the default one where their peers don't act or reproduce well enough to survive, into the one where every species survives and get to compete against each other non-eliminatively, or perhaps even better than that. Take into account all cynical possibilities. I'm not asking about species-level biological strategy. I'm asking about what an individual organism with human-like agency and understanding could do to maximize their lineage's survival probability, even if their conspecifics don't partic…

Data and Society AI 2026-11-15 22:57 UTC Score 39.0 USR-0143-20261115-research-aca-742c7ef7

Tech Companies Make Billions as the Manosphere Fuels Sex Attacks on Younger Girls

Alice Marwick comments about the consequences of the manosphere and online platforms's role in growing incel influence on young people in The Sunday Post (Scottland). The post Tech Companies Make Billions as the Manosphere Fuels Sex Attacks on Younger Girls appeared first on Data & Society .

Trump to meet Xi and Putin in China
Semafor Technology 2026-11-15 22:38 UTC Score 53.0 USR-0094-20261115-global-ai-ne-c5d1132f Full article

Trump to meet Xi and Putin in China

US President Donald Trump is expected to meet Russia’s Vladimir Putin and China’s Xi Jinping in China this week, where he could be upstaged by the controversy around former Nvidia CEO Jensen Huang’s affair with an accused Chinese spy.

LLMs are below chance on TASTE once a shortcut is removed
LessWrong AI 2026-11-15 22:38 UTC Score 83.0 USR-0152-20261115-community-fo-70942552 Full article

LLMs are below chance on TASTE once a shortcut is removed

We found a shortcut in the TASTE benchmark : mech-interp-related proposals are generally scored lower. Considering only the non-interp half, all tested models with ECI > 158 are below 50% on the rubric-only setup, with 8 of 14 significantly below (7 of them OpenAI). In particular, OpenAI models' accuracies average to only 29.6%, and seem to decrease with higher ECI (Pearson r = -0.74), suggesting a systematic bias. The shortcut The TASTE benchmark is a benchmark where models decide which of two AI-safety research proposals is preferred by domain experts [1] . In trying to understand the quality of the benchmark, we realized one recurring theme that the mech-interp-related proposals in there are quite bad. We therefore tested a simple baseline: compare the number of mech-interp related terms (without deduplication) on the two proposals, and pick the one with fewer such terms, coin flip on ties. Surprisingly, this shortcut is able to achieve a 63.0% overall accuracy, beating all the LLMs we tested [2] . The list of mech-interp related terms we use probe, probes, probed, probing, patch, patches, patched, patching, lens, SAE, SAEs, sparse autoencoder, sparse autoencoders, interpretability, attention head, attention heads, residual stream, residual streams, circuit, circuits, circuitry, logit, logits, activations, activation patching, activation steering, activation space, activation vector, activation vectors, activation addition, activation additions, steering vector, steering…

LLMs are weirdly below chance on TASTE benchmark without a shortcut
LessWrong AI 2026-11-15 22:38 UTC Score 83.0 USR-0152-20261115-community-fo-46c5c43c Full article

LLMs are weirdly below chance on TASTE benchmark without a shortcut

We found a shortcut in the TASTE benchmark : mech-interp-related proposals are generally scored lower. Considering only the non-interp half, all tested models with ECI > 158 are below 50% on the rubric-only setup, with 8 of 14 significantly below (7 of them OpenAI). In particular, OpenAI models' accuracies average to only 29.6%, and seem to decrease with higher ECI (Pearson r = -0.74), suggesting a systematic bias. The shortcut The TASTE benchmark is a benchmark where models decide which of two AI-safety research proposals is preferred by domain experts [1] . In trying to understand the quality of the benchmark, we realized one recurring theme that the mech-interp-related proposals in there are quite bad. We therefore tested a simple baseline: compare the number of mech-interp related terms (without deduplication) on the two proposals, and pick the one with fewer such terms, coin flip on ties. Surprisingly, this shortcut is able to achieve a 63.0% overall accuracy, beating all the LLMs we tested [2] . The list of mech-interp related terms we use probe, probes, probed, probing, patch, patches, patched, patching, lens, SAE, SAEs, sparse autoencoder, sparse autoencoders, interpretability, attention head, attention heads, residual stream, residual streams, circuit, circuits, circuitry, logit, logits, activations, activation patching, activation steering, activation space, activation vector, activation vectors, activation addition, activation additions, steering vector, steering…

LLMs are weirdly below chance on TASTE without a shortcut
LessWrong AI 2026-11-15 22:38 UTC Score 80.0 USR-0152-20261115-community-fo-53045649 Full article

LLMs are weirdly below chance on TASTE without a shortcut

We found a shortcut in the TASTE benchmark : mech-interp-related proposals are generally scored lower. On the non-interp half, all tested models with ECI > 158 are below 50% on the rubric-only setup, with 8 of 14 significantly below (7 of them OpenAI). In particular, OpenAI models' scores seem to decrease with higher ECI (r = -0.74), suggesting a systematic bias. The shortcut The TASTE benchmark is a benchmark where models decide which of two AI-safety research proposals is preferred by domain experts [1] . In trying to understand the quality of the benchmark, we realized one recurring theme that the mech-interp-related proposals in there are quite bad. We therefore tested a simple baseline: compare the number of mech-interp related terms (without deduplication) on the two proposals, and pick the one with fewer such terms, coin flip on ties. Surprisingly, this shortcut is able to achieve a 63.0% overall accuracy, beating all the LLMs we tested [2] . The list of mech-interp related terms we use probe, probes, probed, probing, patch, patches, patched, patching, lens, SAE, SAEs, sparse autoencoder, sparse autoencoders, interpretability, attention head, attention heads, residual stream, residual streams, circuit, circuits, circuitry, logit, logits, activations, activation patching, activation steering, activation space, activation vector, activation vectors, activation addition, activation additions, steering vector, steering vectors, integrated gradients Analyzing the other hal…

SpaceX makes history by catching spacecraft
Semafor Technology 2026-11-15 22:36 UTC Score 57.0 USR-0094-20261115-global-ai-ne-af0f8d23 Full article

SpaceX makes history by catching spacecraft

SpaceX caught its Starship spacecraft in giant mechanical arms for the first time Saturday, clearing another obstacle en route to Elon Musk’s vision of making crewed space launches as regular as air travel.

Anthropic IPO set to define AI trade
Semafor Technology 2026-11-15 22:36 UTC Score 58.0 USR-0094-20261115-global-ai-ne-3c749a04 Full article

Anthropic IPO set to define AI trade

Anthropic’s $1.4 trillion IPO on Thursday sets up a defining test for an AI trade buffeted by market turbulence, mounting political backlash, and a crisis at Nvidia.

Faulty router, aging infrastructure blamed in US flight stoppage
Semafor Technology 2026-11-15 22:35 UTC Score 49.0 USR-0094-20261115-global-ai-ne-726e7ef9 Full article

Faulty router, aging infrastructure blamed in US flight stoppage

A single faulty router may have triggered the nationwide ground stop ordered by the Federal Aviation Administration on Friday, underscoring the dangers posed by the US’ outmoded technological infrastructure.

Iran attacks US military base in Bahrain as talks stall
Semafor Technology 2026-11-15 22:34 UTC Score 57.0 USR-0094-20261115-global-ai-ne-1e1c0cfe Full article

Iran attacks US military base in Bahrain as talks stall

Iran launched drone attacks against a US military base in Bahrain on Sunday, adding to mounting tensions even as Washington signals a reluctance to escalate.

Waluigi World
LessWrong AI 2026-11-15 22:16 UTC Score 50.0 USR-0152-20261115-community-fo-c9db0740 Full article

Waluigi World

She was like the pause in the room but didn’t know it yet — pregnant. Tatiana lit a cigarette. She had volunteered to speak first. “Spring was late, and so was my period. Something has been going on recently. I invited over these two I met at a bookstore, a guy and a girl. The guy – as soon as he looked at me, I froze. My vulva was as cold as his heart. I was sitting down. She walked over to me. She came right up to me and pushed my head into her thighs. Her vagina was like a womb.” Waluigi nodded wisely. Though it was his first time running a support group, Waluigi was a natural. Women, he felt, had always seemed comfortable in his presence. Perhaps he would consider a career as an OBGYN. Or maybe… “We exchanged body fluids and ideas. She said she filled her god-shaped hole with Russian literature and American semen. That’s why she picked me – I’m writing my dissertation on Dostoevsky. She left me Dostoevsky’s Demons and for two and a half weeks. Eros kept me warm that night. Eros is that homeless guy that lives on twelfth street. “The next night, the guy came back. ‘So you study the big D, eh?’ His joke was lame, like a person with a limp. I asked him about himself. His three favorite things in life were making money, love, and off with another man’s money. I told him about Eros. He passed judgement and then out. I guess his heart was as hard as his cock.” Something stirred within Waluigi. “When were you expecting your period?” “Let me think… Oh. Oh, I guess it should have…

The Verge AI 2026-11-15 21:49 UTC Score 52.0 AI-016-20261115-global-ai-ne-e2c4ca29 Full article

Listen to this: Armand Hammer’s Snakes & Ladders

I open with this caveat: I haven't had a ton of time to digest the surprise Armand Hammer record that dropped this week. Billy Woods and Elucid rarely miss, though, and my initial reaction is that Snakes & Ladders ranks among their best. It doesn't have the seamless throughline production-wise that their collabs with Alchemist […]

What Futures Do AI Models Want? An Exploratory Character Evaluation of AI for AI Strategy
LessWrong AI 2026-11-15 21:47 UTC Score 74.0 USR-0152-20261115-community-fo-9acb118a Full article

What Futures Do AI Models Want? An Exploratory Character Evaluation of AI for AI Strategy

(This is cross-posted from my blog ) Epistemic Status: Evaluation represents approximately 2 weeks of work. This is mostly exploratory evaluation, and the code for evaluation was made using LLMs. “Before I draw nearer to that stone to which you point, answer me one question. Are these the shadows of the things that will be, or are they shadows of things that may be only?” ~ Charles Dickens Models may soon be involved in some of the most important decisions humanity makes about its long-term future. These include choices about the proliferation of digital minds, tradeoffs between extinction risk and loss of human control, whether to race toward or pause AGI development, and which interventions to prioritize for humanity’s deep future. While there may not be clear answers to many of these questions, some choices are better. It is therefore important to understand models’ preferences and character, both to evaluate the decisions they might make and to understand how they could influence our own. Even without fully delegating these decisions to AI systems, they may have substantial influence through advice, research, and recommendations (see this study from Anthropic ). I’ve developed a small exploratory evaluation of models’ preferences about AI strategy and the long-term future, including decisions about running a frontier AI lab and allocating philanthropic funding. It combines my own questions, with wording refined by Astra, and adaptations of questions from Hard Cases for A…

Vulcan rocket fails, destroying 40 Amazon Leo satellites
The Verge AI 2026-11-15 21:05 UTC Score 57.0 AI-016-20261115-global-ai-ne-f4d10e35 Full article

Vulcan rocket fails, destroying 40 Amazon Leo satellites

United Launch Alliance (ULA) suffered what is reportedly its first mission failure in its 20-year history, failing to deliver 40 Amazon Leo satellites to orbit. Leo is Amazon's competitor to SpaceX's Starlink, but it has struggled to get off the ground so far, putting only a small portion of the necessary satellites to challenge Starlink […]

The Verge AI 2026-11-15 20:43 UTC Score 52.0 AI-016-20261115-global-ai-ne-5243c0b6 Full article

Jensen Huang might be plotting a return and more of the latest Nvidia drama

The drama around Nvidia co-founder and former CEO Jensen Huang is evolving quickly. The latest edition of our Stepback newsletter should get you up to speed, but there have already been a number of new developments since this morning. That includes rumors that Jensen Huang might be trying to find a path back to CEO, […]

젠슨 황, 엔비디아 CEO 복귀 논의…측근들 백악관 설득 추진
Korea AI Times 2026-11-15 20:32 UTC Score 35.0 USR-0048-20261115-global-ai-ne-51477d48 Full article

젠슨 황, 엔비디아 CEO 복귀 논의…측근들 백악관 설득 추진

젠슨 황 전 엔비디아 CEO가 사임한 지 며칠 만에 측근들과 복귀 가능성을 논의하기 시작한 것으로 알려졌다. 실리콘밸리 주요 경영진은 복귀를 위한 조직적인 로비에 나서고 백악관의 지원까지 확보할 계획이다.파이낸셜타임스(FT)는 15일(현지시간) 복수의 소식통을 인용해 황 전 CEO가 측근들과 엔비디아 CEO 복귀 가능성을 논의하고 있다고 보도했다.황의 측근 한 명은 실리콘밸리 주요 경영진으로 구성된 \'워룸(war room)\'을 조직해 복귀 전략을 조율하고, 엔비디아 이사회를 설득하는 동시에 백악관의 지지도 확보할 계획이라고 밝혔다.

Where are the adults in AI safety?
LessWrong AI 2026-11-15 20:10 UTC Score 56.0 USR-0152-20261115-community-fo-f695268a Full article

Where are the adults in AI safety?

The following is based on a true story (not mine). Let’s say you’ve been bombarded with enough AI doom that you decide to do something about it! You watched Artificial and don’t trust the frontier labs to steer us somewhere safe on their own, so you want to work on the problem from the outside. You’re a little fuzzy on the acronyms - AGI, RSI, TAI, SVU - but by God you’re going to figure it out. So, you Google “How to get into AI safety.” And it’s encouraging! There are articles about how AI safety needs founders. There are calls for technical people to switch careers. There are career transition grants, which sound like money to stop doing your meaningless and/or actively destructive job and start saving the world. “Great, that’s me!” you think, and so you apply. Within 24 hours you get a rejection email in your inbox: “Hey, we’re so happy you want to save the world. It’s the most important issue of our generation! But we don’t think you can help. Why don’t you try taking my course or reading my LessWrong articles or listen to my podcast or follow me on twitter first? Thanks!” Huh. Maybe it’s because you used an AI to edit your answers. You’d think this field, of all fields, would have made its peace with that. Maybe you’re just not that good. Fine. So, you take the seemingly automated email’s advice and sign up for a course from the color-shape company, to learn and meet like-minded people. A few hours after signing up - err, applying? they don’t just let you sign up - ano…

Congress wants to tighten export rules on AI chips after Nvidia spy scandal
The Verge AI 2026-11-15 19:15 UTC Score 63.0 AI-016-20261115-global-ai-ne-bc766627 Full article

Congress wants to tighten export rules on AI chips after Nvidia spy scandal

Unsurprisingly, following the scandal involving now-former Nvidia CEO Jensen Huang and a suspected Chinese spy, Congress is looking to further restrict exports of AI chips. Both Democrats and Republicans have voiced support for various measures and bills that would tighten rules and have demanded that Huang and other Nvidia leaders testify before Congress. Senate Minority […]

Giving Up on Saving the World
LessWrong AI 2026-11-15 19:00 UTC Score 53.0 USR-0152-20261115-community-fo-b08a710d Full article

Giving Up on Saving the World

When I was 19, I decided I was going to save the world. Now, 25 years later, I’ve decided to give up. To explain why and what giving up means, we first need to go back to how I ever got into the world-saving business. As a kid, I loved computers, and I started teaching myself to program them when I was 8. By 16 I could write real programs that did interesting things, and it was also around this time I got into Linux. And since I was into Linux, that meant I read Slashdot . If you weren’t into Linux in the late 90s, you might not even know what Slashdot was. It still exists, but it’s a shell of its former self. Back then it was a vibrant, community-driven news site that in many important ways felt to me like LessWrong feels now. And it was from Slashdot that I learned about nanotech and grey goo, which led me to the Foresight Institute, which led me to the Extropians, which ultimately led to me being a very early member of the SL4 mailing list. And, again, if you’re not familiar with SL4, it’s approximately how all this AI safety and rationalist stuff got started. And so it was sometime over the summer of 2001, discussing Friendly AI with Eliezer Yudkowsky on IRC, that I decided to orient my life around saving the world from the risks of superintelligent AI. The only problem was, I wasn’t up to the task yet. In fact, nobody was. In that first decade I, along with everyone else, was just trying to figure out how to think well enough to be able to think usefully about AI at all…

Easier To Be Good
LessWrong AI 2026-11-15 18:34 UTC Score 47.0 USR-0152-20261115-community-fo-6a159e52 Full article

Easier To Be Good

I do not consider myself a good person. I am lazy and frivolous and inconsiderate, I care too much about intelligence and status, I am a snob and a hypocrite, and I do not call my mother enough. I don't mean this to read as self-hatred, nor a show of false modesty. Among all the people I know there is only a scant handful of people I would consider genuinely good, and so it is unremarkable to me that I wouldn't find myself amongst their number. You do not need to be good to desire more goodness in the world. It is good for everyone when things are good, and when there is more of it — more thoughtfulness, more generosity, more coordination, more care, more beauty, more clarity, more positive sum trades, more sophistication and elegance in our solutions and resolutions. Increasing the amount of goodness in the world is an urgent project. The world really sucks right now, and even in the wealthiest enclaves of the imperial core there is an immense amount of suffering . To reach the healed world would be a massive complicated undertaking. It is such a massive complicated undertaking that it is not enough for only those who are intrinsically good to work on it. There simply aren’t enough of them. If it is desirable (as I contend) for morally mediocre people like me to work on the project, it would not be enough to appeal to their innate goodness, for they have not that much of it to be appealed to. You will need to offer me things that are more base, like status and money and the…

Bayesian Games with Nukes
LessWrong AI 2026-11-15 18:22 UTC Score 47.0 USR-0152-20261115-community-fo-4ce36a6f Full article

Bayesian Games with Nukes

How we all learned to stop worrying and love the bomb... for a while Listen to this post on the Non-Zero-Sum James podcast Nearing the end of 1961 the Soviet Union tested a 50-megaton hydrogen bomb equivalent to 3800 Hiroshima bombs. At the same time an intelligence officer by the name of Juanita Moody and her team became aware of Soviet ships delivering large cargo, under cover of darkness in Cuba—what if such a bomb was being delivered to the doorstep of the US? A surprise attack of that magnitude, would devastate the US, and potentially destroy its ability to retaliate, jeopardising the offensive standoff between the Soviets and the US that would come to be known as... M.A.D: An Easy (Albeit High-Stakes) Problem One of the strangest dynamics in geopolitical history has to be that of Mutually Assured Destruction (M.A.D.), it is almost as crazy as its name suggests. "Two superpowers, the U.S. and U.S.S.R., represent such enormous destructive potentials as to afford little chance of a purely passive equilibrium" —John Von Neumann The father of game theory, John Von Neumann correctly identified that a passive approach to nuclear war would not result in a peaceful equilibrium, instead a mutually aggressive stance counter-intuitively lead to the long-running stand off that was the Cold War, and except for a few close calls, humanity found an unlikely saviour. Miraculously, mutual aggression with apocalyptic weaponry has shepherded the world through the most peaceful period in r…

Alignment 'train/test splits' should be sacred
LessWrong AI 2026-11-15 18:22 UTC Score 78.0 USR-0152-20261115-community-fo-fa861d15 Full article

Alignment 'train/test splits' should be sacred

Tl;dr: Alignment techniques can be used for training alignment or for testing it, but they cannot be used for both! As we build out the set of alignment techniques available to us, we should enforce that some are judiciously reserved only for evaluating alignment. See also: Validating against a misalignment detector is very different to training against one . Here is a guiding principle regarding alignment-minded AI training: whatever you do, you must maximally avoid teaching the model to take misaligned actions which are invisible to the user. In fact, you should be shamed if you use training techniques which encourage covert misalignment. Here is a separate, guiding principle of machine learning: whatever you do, you must maximally avoid training the model on the entire dataset rather than holding out a test split. In fact, you should be shamed if you report results which are obtained by training on the entire dataset rather than on the train split. I want to argue that these are the same principle. Training against all of your alignment assessment techniques leads to the same failure as training on your validation data in traditional machine learning: you potentially get better performance, but you sacrifice all ability to verify your model’s behavior. Consider two types of alignment techniques: Some techniques sculpt the model’s alignment. This includes techniques such as: Steering and other inference-time interventions SFT generalization techniques (midtraining, pretrai…

The Guardian view on civil service reform: AI cannot conjure a capable state on its own | Editorial
The Guardian AI 2026-11-15 17:25 UTC Score 45.0 AI-021-20261115-global-ai-ne-ed64519c Full article

The Guardian view on civil service reform: AI cannot conjure a capable state on its own | Editorial

Whitehall needs to change, but more technology isn’t enough. Opening routes to leadership would bring skills that are too often overlooked The cabinet secretary, Antonia Romeo, called for an overhaul of a “ reactive and risk-averse ” civil service in her speech last Tuesday. Ministers planned a “radical shift” of powers, resources and accountability to regional and local government, which she said would leave Whitehall “leaner and more strategic”. As 80% of civil servants already work outside London, devolution promises to alter what central government does, not simply where it operates. Reinvention is in the civil service’s blood – ever since it was founded as a permanent, impartial institution in 1854. Even departmental boundaries, Dame Antonia said, are on notice, asking do they “remain optimal in an era of cross‑cutting delivery?” This must be set against the civil service’s relative decline: it is ranked joint sixth internationally on one measure of public administration. To improve, Dame Antonia says the civil service must embrace AI. The efficacy of public service, she claimed, no longer has to be constrained by the limits of “ human bandwidth ”. That is an extraordinary claim. Technology can help speed up identifying someone at risk of homelessness – but it does not supply a home. Better administration and adequate resources need each other. Continue reading...