AI News

Banger paper from CMU on harness learning.

(bookmark it)

Also, pay attention to this important new AI engineering skill of improving agents by editing their harness code instead of their weights.

Seeing a huge shift towards this.

The authors train a proposer model with RL to read a task, the cu…

Taiwanese semiconductor giant TSMC is considering collaborating with Elon Musk's Terafab project to establish advanced end-to-end semiconductor manufacturing in the United States. According to an exclusive report from Tim Culpan's Culpium newsletter, later confirmed by Elon Musk on X, th

This week’s list of the largest U.S. startup funding rounds was pretty much all about AI. That includes the biggest financing — a $1 billion round for Instinct, a developer of AI assistants for everyday tasks — as well as most of the rest of the Top 10.

I dati dell’Osservatorio del PoliMi: le competenze IA raddoppiano negli annunci italiani. Ma la corsa ora riguarda marketing, risorse umane e legal: il vero mo…

Invitalia autorizza l’investimento, il cantiere potrebbe partire già da metà 2027: previsti 1600 posti di lavoro

You can now select images or listings for clothes that appear in the chatbot, tap “Try on,” and then upload images of yourself to see what you’d look like wearing the apparel. Sound familiar? That’s because Google introduced a near-identical AI try on feature over a year ago.
[Image: In this example ChatGPT’s feature even dyed and styled her hair, presumably without being asked to, so I’m sure we can trust the try on results… https://platform.theverge.com/wp-content/uploads/sites/2/2026/10/Virtual-Try-On.jpg?quality=90&strip=all]

Exynos 2700 is reportedly in mass production, a tipster claims, with 10% higher volume than Exynos 2600 and 10% faster multi-core speeds than Snapdragon

The zero-water architectures improve efficiency and can redirect up to 22 MW of electrical capacity to AI workloads.

3/ @zainhas and @parthsareen (Agents at @ollama) took the breakout stage on evaluating open models against production traffic (not just benchmarks).

Model collapse arises when generative models are trained on synthetic data produced by earlier models. The phenomenon has attracted considerable attention because of its societal and technical implications. However, previous studies have reached seemingly contradictory conclusions: replacing real data with synthetic data causes collapse (Shumailov et al.), yet accumulating real data alongside synthetic data can prevent it. For diffusion models, we study an intermediate regime typical of finite-budget pipelines: all past datasets and the real data are kept, but each new model is trained on a fixed-size sample from this growing pool, so the real fraction vanishes without any data being removed. Experiments on a 2D spiral dataset as well as the image benchmarks (MNIST, Fashion-MNIST, and CIFAR-10) show that replacement protocol degrades dataset rapidly as in the literature, whereas the fixed budget degrades only partially, sparing some features. A linear-response model of the multi-generational parameter dynamics, analyzed by stochastic recursion, confirms that the two protocols differ: some features will be fragile and lost within a few generations for both protocols, while some will be robust and preserved over practically unbounded horizons under the fixed budget protocol.

Recent years have witnessed the rapid adoption of reinforcement learning (RL) in large language model (LLM) post-training, with substantial gains in mathematical reasoning and code generation. In practical systems, however, policy updates and differences between rollout and training engines can make sampled responses off-policy. Sequence-level masking addresses this mismatch by deciding whether an entire response should contribute to optimization. A common masking rule uses the length-normalized geometric mean of sampled token probability ratios. Its signed log-ratios can cancel across positions, concealing substantial bidirectional policy drift. We propose \emph{Cancellation-Aware Response Masking} (CARM), a sequence-level mask that takes the absolute value of each token log-ratio before averaging, preventing opposing probability changes from canceling. We prove that accepted responses satisfy a joint bound on the fraction of sampled-token ratios outside a prescribed band and their mean log-distance beyond its boundaries. Experiments on mathematical reasoning and code generation show that CARM improves mean@16 averaged over AIME 2024/2025/2026 and BeyondAIME by up to $3.13$ percentage points over geometric-mean masking, and increases average pass@1 across four code benchmarks by $2.88$ points over the strongest evaluated baseline. These findings support CARM as a theoretically grounded and effective method for response-level off-policy control in LLM reinforcement learning.

Startup funding is relatively quiet this morning, but one deal stands out. PaleBlueDot AI raised $200 million in Series C funding at a $3.2 billion valuation, putting AI infrastructure firmly at the center of today’s funding activity as the company expands the GPU capacity behind large-scale AI workloads. Beyond that blockbuster round, investors are placing

Per aiutare i valdostani a maneggiare gli strumenti della nuova rivoluzione tecnologica disponibili 24 ore di lezioni

Ispirato all’esperienza di America.gov, il progetto permette di chiedere in linguaggio naturale informazioni per orientarsi tra servizi e portali della PA. Il …

Suno is branching out from the world of AI music, launching a new feature that generates spoken voices based on scripts or prompted descriptions.

In mobile robots, software, safety, and fleet management are becoming just as important as the robot itself, says an OMRON product manager.

Astribot, a Shenzhen, China-based physical AI company, showcased its T1 humanoid robot to North America for the first time at IROS 2026, which took place from September 27 to October 1 at the David…

We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:

Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance d…

Join the discussion on this paper page

Asynchronous reinforcement learning (RL) improves the efficiency of large language model post-training but introduces stale rollouts generated by earlier policies. Theoretical understanding of how this staleness affects convergence and how to mitigate its impact remains limited. We derive a convergence bound for GRPO-style algorithms that explicitly characterizes the tradeoff between the gradient estimator's second moment and bias. For trajectory-level importance-weighted estimators, our analysis shows that once the second moment is uniformly controlled, delay enters the bound through the bias introduced by clipping or rescaling. Guided by this insight, we propose a novel group mass capping GRPO (GMC-GRPO) method, which minimizes a ratio-based bias bound within a class of weighted estimators sharing a common second-moment guarantee. We establish convergence guarantees for asynchronous GMC-GRPO and show that, compared with TIC-GRPO, it improves the threshold dependence of the fourth-order delay term from $O(ε^{-4})$ to $O(ε^{-2})$ as $ε\to0$, where $1+ε$ is the ratio threshold. Under local policy overlap, the delay-dependent term decreases as $G^{-2/5}$ after tuning the step size, where $G$ is the group size. For fixed behavior and current policies, the bias introduced by group rescaling also vanishes as $G\to\infty$, whereas the bias from trajectory-wise clipping can persist. Experiments across Qwen3 models and reasoning benchmarks demonstrate improved robustness to stale rollouts, with GMC-GRPO achieving the best performance among stable baselines under large rollout delays.