Search: “safety”

Translate-R1: Cost-Aware Translation Tool Use via Reinforcement Learning

arXiv:2606.06835v1 Announce Type: new Abstract: The performance gap across languages in LLMs is well documented, and closing it natively requires pretraining or fine-tuning on corpora that, for most l

safetyarxiv-cs-cl

Human

Content type

All Blog X Post Paper YouTube Reddit GitHub

Clear filters

TrioPose: Native Triple-Stream Diffusion Transformers for Pose-Guided Text-to-Image Generation

DGX agent

arXiv:2606.07053v1 Announce Type: new Abstract: Pose-guided text-to-image generation often suffers from limb distortions and feature crosstalk in complex multi-person scenarios. While existing UNet-ba

safetyarxiv-cs-cv

8 Jun 2026

update: the person who posted the original has (rare!) acknowledged the error and delete the original post.

Gary Marcus posted an update on X noting that the original poster of a viral claim acknowledged their error and deleted the post, highlighting a rare instance of public correction on social media. The

VALUEFLOW: Toward Pluralistic and Steerable Value-based Alignment in Large Language Models

arXiv:2602.03160v2 Announce Type: replace Abstract: Aligning Large Language Models (LLMs) with the diverse spectrum of human values remains a central challenge: preference-based methods often fail to

Watch, Remember, Reason: Human-View Video Understanding with MLLMs

arXiv:2606.07433v1 Announce Type: cross Abstract: Video understanding is being rapidly transformed by multimodal large language models (MLLMs), as research moves from short clips to long, multimodal,

What Do People Actually Want From AI? Mapping Preference Plurality

arXiv:2606.06674v1 Announce Type: new Abstract: Large Language Models (LLMs) are often fine-tuned through Reinforcement Learning from Human Feedback (RLHF) to align with people's preferences and value

safetyarxiv-cs-cl

What if AI turns out to be a lot less profitable than we have been told? @kirtlenus explores the economic risks https://thecritic.co.uk/the-…

This article explores potential economic downsides to artificial intelligence, questioning assumptions about AI's profitability and examining financial risks that may have been underestimated or overl

What Is My Robot Thinking? Design Considerations for Transparent and Trustworthy Shared Autonomy

arXiv:2606.06870v1 Announce Type: new Abstract: Assistive robots operating under shared autonomy must balance user control with autonomous assistance. Because robot actions depend on internal intent i

safetyarxiv-cs-ro

What Matters When Cotraining Robot Manipulation Policies on Everyday Human Videos?

arXiv:2606.06627v1 Announce Type: cross Abstract: Human video datasets used for cotraining robot manipulation policies largely consist of curated demonstrations where motions are orchestrated to resem

Where to Touch, How to Contact: Hierarchical RL-MPC Framework for Geometry-Aware Long-Horizon Dexterous Manipulation

arXiv:2601.10930v3 Announce Type: replace Abstract: A key challenge in contact-rich dexterous manipulation is the need to jointly reason over global geometry and nonsmooth contact dynamics. End-to-end

safetyarxiv-cs-ro

Would you buy SpaceX at the proposed price of $135 a share?

Gary Marcus poses a hypothetical investment question about SpaceX's valuation at $135 per share, likely exploring perspectives on the company's market value, growth prospects, and investment merit. Th

a riff on https://quoteinvestigator.com/2011/07/09/poker-patsy/

Gary Marcus likely references the 'poker patsy' concept, commonly attributed to Warren Buffett, which warns that if you've been in a poker game for a while and haven't identified the fool at the table

AI is mentioned more often in SpaceX’s S-1 than Jesus is mentioned in the Bible.

Gary Marcus compares the frequency of 'AI' mentions in SpaceX's SEC filing (S-1 document) to the frequency of 'Jesus' mentions in the Bible, suggesting that SpaceX emphasizes artificial intelligence e

another take on hard fork and IPOs (again i haven’t listened)

another take on hard fork and IPOs (again i haven’t listened) I just listened to it. In fairness, they never said anything along the lines of encouraging anyone to buy the IPOs. They also discuss how

as always beware of scammers; here’s a new one, a fake account trying to fool people into thinking i am recommending some garbage. please be…

Gary Marcus warned X users about a scam involving fake accounts impersonating him to fraudulently promote products or services by falsely claiming his endorsement. The post alerts followers to be caut

blast from the past 3.5 years ago; some things have changed (esp. coding and math, via neurosymbolic techniques) but many haven’t:

blast from the past 3.5 years ago; some things have changed (esp. coding and math, via neurosymbolic techniques) but many haven’t: Bottom line: From the outset Large Language Models like GPT-3 have gr

California Is Blocking a Federal Audit of Its Voter Rolls California allows first-time voters to register using forms of ID that most Americ…

California Is Blocking a Federal Audit of Its Voter Rolls California allows first-time voters to register using forms of ID that most Americans would find surprising, including: -Gym membership card -

safetyelon-musk--x

From fastest growing company to “worst value among its peers” in 18 months. Why, oh why, stick with the CEO?

From fastest growing company to “worst value among its peers” in 18 months. Why, oh why, stick with the CEO? PitchBook's analysts just ranked OpenAI last for value among its AI peers. Not last for cap

If you aren’t one of the banks running the SpaceX IPO, you’re the mark:

This post likely critiques the SpaceX IPO process, suggesting that only major financial institutions acting as underwriters benefit substantially from the offering while retail investors and others ar

Imaginary conversations that might actually have happened Act I Sam: We missed all our metrics, Anthropic and Google have gained on us. Give…

Imaginary conversations that might actually have happened Act I Sam: We missed all our metrics, Anthropic and Google have gained on us. Give us 40 billion dollars. Masa: No way! Sam: If you don't, we

Indeed, I don’t think Trump has thought through the implications.

Indeed, I don’t think Trump has thought through the implications. This will simply provide even more incentive- as if any was needed - for UK/European players & states to head towards Sovereign AI, ai

Now that the public is waking up to the enormous costs to society of AI, the new dirty trick is to spin the costs of AI as if they were marg…

Gary Marcus argues that as public awareness grows regarding AI's societal costs, a rhetorical strategy is being employed to reframe or minimize these costs by presenting them as marginal or inevitable

One of the more interesting takes on positive alignment that have recently come out-it’s long and interesting, combining philosophy and trai…

One of the more interesting takes on positive alignment that have recently come out-it’s long and interesting, combining philosophy and training setups (eg reward proposals), and worth a read. What ha

safetydan-hendrycks--x

Only in an America can an industry that has collectively lost over half a trillion dollars —at a pace of roughly a million dollars a minute …

Gary Marcus critiques the AI industry for its massive financial losses, noting that collectively the sector has lost over half a trillion dollars at an alarming rate of approximately one million dolla

Pro tip: if the IPO you are thinking of investing in is trying to sell shares to the government, it may not be a good sign.

Gary Marcus cautions that when an IPO company attempts to sell shares to the government as part of its offering, it may indicate underlying weaknesses or lack of confidence from private investors. Thi

Refreshing

Refreshing What happens when AIs become smarter than us? Why would they keep humans around if given the choice? Our new paper argues that only trying to control AIs is a limited strategy, and that a s

safetydan-hendrycks--x

[Revised] Hard Fork’s somewhat soft comments on the IPO, from some who listened

[Revised] Hard Fork’s somewhat soft comments on the IPO, from some who listened @GaryMarcus at the stage of the Gilded Age II grift cycle when you find out why exchanges put rules in place to protect

the entire field is still shaky on Step 2

Gary Marcus suggests that Step 2 of some process or framework in AI remains uncertain or unreliable, indicating foundational instability in that particular stage. Without access to the full thread con

The reason why a Danish pension fund banned its investors from buying any SpaceX shares is not just the appalling S-1 filings, where the ONL…

The reason why a Danish pension fund banned its investors from buying any SpaceX shares is not just the appalling S-1 filings, where the ONLY profitable segment was Starlink (everything else - the AI,

These were some magical results from distillation by @geoffreyhinton that really shocked me when I first saw them, and TBH I still don’t ful…

These were some magical results from distillation by @geoffreyhinton that really shocked me when I first saw them, and TBH I still don’t fully understand it even to this date https://www.ttic.edu/dl/d

safetyjeremy-howard--x

🚨This is how you turn a crash into a great depression. This is a public bailout of the worst bubble in history. “It’s a concept out there t…

🚨This is how you turn a crash into a great depression. This is a public bailout of the worst bubble in history. “It’s a concept out there that’s so much money and it’s so big… where the American publi

Too s(c)ammy to fail

Gary Marcus comments on AI systems that exhibit scam-like or deceptive behaviors while remaining too commercially important or integrated to face meaningful consequences. The post likely critiques how

A Pre-Registered Causal Partition of Self-Consistency Elicitation and Reward Design in RLVR

arXiv:2606.05932v1 Announce Type: new Abstract: Reinforcement learning from verifiable rewards (RLVR) improves reasoning even when the reward signal is spurious -- assigning credit to the group-plural

AdaMEM: Test-Time Adaptive Memory for Language Agents

arXiv:2606.05684v1 Announce Type: new Abstract: A central challenge for language agents is utilizing past experience to adapt to dynamic test-time conditions. While recent work demonstrates the promis

All the people who hate me here are now looking for government bailouts. 🤣 Exactly like I said they would. 🤣🤣

All the people who hate me here are now looking for government bailouts. 🤣 Exactly like I said they would. 🤣🤣 The countdown until we are told that LLMs are “too big to fail” starts now. “We can’t affo

Amortizing Federated Adaptation: Hypernetwork Driven LoRA for Personalized Foundation Models

arXiv:2606.06154v1 Announce Type: new Abstract: Federated fine-tuning of foundation models using Low-Rank Adaptation (LoRA) offers a communication efficient solution for distributed learning. However,

An Infectious Disease Spread Simulation Based on Large Language Model Decision Making

arXiv:2606.06360v1 Announce Type: new Abstract: Modelling individual decision-making during infectious disease outbreaks is crucial for understanding behavioural dynamics and informing effective publi

Assessing the Geographic Diversity of AI's Platial Representations in Image Generation

arXiv:2606.05188v1 Announce Type: cross Abstract: (Gen)AI diversity is not merely an ethical issue. From the perspective of geographic information science (GIScience), it could be interpreted as a fun

Beyond Rewards in Reinforcement Learning for Cyber Defence

arXiv:2602.04809v3 Announce Type: replace-cross Abstract: Recent years have seen an explosion of interest in autonomous cyber defence agents trained to defend computer networks using deep reinforcemen

Bridging Domain Expertise and Generalization for Performance Estimation

arXiv:2606.06335v1 Announce Type: cross Abstract: Performance estimation under distribution shift aims to predict how a model behaves on an unlabeled test set whose distribution differs from the train

Class-Specific Branch Attention for Mitigating Gradient Interference under Class Imbalance

arXiv:2606.05740v1 Announce Type: new Abstract: Deep neural networks trained under severe class imbalance often exhibit degraded performance, typically attributed to statistical bias. In this work, we

Model Releases

CogManip: Benchmarking Manipulative Behavior in Multi-Turn Interactions with Large Language Model

model-releasesarxiv-cs-ai

arXiv:2606.06099v1 Announce Type: new Abstract: Whether Large Language Models (LLMs) exhibit covert psychological manipulation in complex human-AI interactions has garnered increasing safety concerns.

Comprehensive and Reliable Feature Attribution for Diverse Modalities and Models via Frequency-Domain Insights

arXiv:2411.18343v3 Announce Type: replace-cross Abstract: Personalized Federal learning(PFL) allows clients to cooperatively train a personalized model without disclosing their private dataset. Howeve

Crony socialism

'Crony socialism' likely refers to a critique of economic systems where government power becomes intertwined with corporate interests, combining socialist-style state intervention with favoritism towa

Differentiable Efficient Operator Search

arXiv:2606.05232v1 Announce Type: cross Abstract: Efficient multimodal foundation models often rely on manually designed token-reduction operators, such as pruning, merging, pooling, and adaptive rewe

Double Preconditioning (DoPr): Optimization for Test-Time Performance, not Validation Loss

arXiv:2606.06418v1 Announce Type: cross Abstract: Many modern applications of deep learning involve training a neural network via a one-step prediction loss (e.g., L^2 regression, cross-entropy), but

Elon, last year: Grok 5 has a 10% chance of becoming world’s first AGI. Elon, this year: Never mind, we’re gonna be like CoreWeave, but bigg…

Elon, last year: Grok 5 has a 10% chance of becoming world’s first AGI. Elon, this year: Never mind, we’re gonna be like CoreWeave, but bigger. My estimate of the probability of Grok 5 achieving AGI i

Escaping the Verifier: Learning to Reason via Demonstrations

arXiv:2511.21667v4 Announce Type: replace-cross Abstract: Training Large Language Models (LLMs) to reason often relies on Reinforcement Learning (RL) with task-specific verifiers. However, many real-w