FERPO preprint reports faster actor updates than REPPO
FERPO replaces critic action derivatives with a forward-KL fit to an entropy- and KL-regularized target distribution; its reported gains rest on unreviewed experiments....
Read article →
Adrian Cole
Infrastructure Correspondent
Chips & Infrastructure
Adrian Cole follows the infrastructure behind the industry: chips, data centers, power and the economics of compute. His coverage asks what systems cost to build and operate, where the bottlenecks sit and which performance claims survive a like-for-like comparison.
Articles by Adrian
905 published articles · Showing 61–80
Preprint Examines Geometry of Inference in Transformer Residual Streams
The preprint shows that transformer representations become more specific to their final output through directional alignment rather than Euclidean distance, offering a geometric explanation for inference convergence....
Read article →AlphaDLM Training Objective Aims for Better Accuracy-Computation Trade-off, Preprint Claims
The approach targets a weakness of factorized diffusion models: cross-entropy training fits conditional token marginals, but parallel token generation requires consistent joint predictions....
Read article →Inverse Distillation Unlearning Combines Data Removal and Model Distillation
The approach could reduce the inference cost of diffusion and flow-matching models, requiring only a pretrained teacher and examples from the data to be forgotten, with no access to the original retained training data, the abstract states. arXiv CS....
Read article →Preprint Says Top-k Accuracy Metric Has 'Blind Spot' in RCA Benchmarks
Root cause analysis is critical for preventing safety incidents and costly downtime in complex monitored systems, the authors note, but the standard top-k accuracy metric can obscure whether a method retrieves the correct cause or simply ranks it poorly....
Read article →Suno Launches AI Spoken Word Generator with Music
The launch expands Suno beyond its core AI music service into the competitive AI-generated speech market, where platforms such as ElevenLabs, Adobe, and DeepMind have been active for years, The Verge noted....
Read article →Preprint Probes When Recurrence Aids Looped Language Models
The study aims to clarify when additional recurrence improves performance and how design choices dictate effectiveness, a question central to scaling inference compute without expanding model size....
Read article →AWS Shows Compliance Pattern for Lease Scanning with Amazon Quick
The pattern, called Adjudicated Query, aims to deliver provable completeness and defensibility in regulatory checks that can be challenged in audits or litigation, the post states....
Read article →Denver Club Saves 10-15 Hours a Week with ChatGPT Work, OpenAI Says
OpenAI published a case study describing how a Denver social club cut administrative task times from days to hours using ChatGPT Work, with no independent corroboration provided....
Read article →Boston Dynamics unveils redesigned Atlas hand with 13 degrees of freedom
Boston Dynamics trades anthropomorphic looks for ruggedness and low-cost mass production in its new Atlas hand, aiming to scale humanoid robotics from research to real-world deployment....
Read article →Preprint Proposes SMAT Merge-Aware Training With Under 2% Overhead
A new preprint describes a merge-aware training method that improves merged-model performance while adding less than 2% overhead to standard fine-tuning, potentially making model merging more practical....
Read article →Allen AI Releases Olmo-core 3, MoE Training Stack Scaling to 1T+ Parameters
The open-source release from Allen AI reduces the overhead of large expert pools; a 47B model with 128 experts lost under 5% throughput compared to an 8-expert version. It positions the Olmo ecosystem for an upcoming MoE model....
Read article →NVIDIA Claims AI Factories Boost ROI Through Higher Throughput, Longer Life
NVIDIA’s ROI pitch for AI factories rests on three claims: higher per-watt productivity, longer hardware earning life, and the ability to run varied workloads. The company provided no independent validation of the cited performance gains or economic assumptions....
Read article →Preprint Proposes Adversarial Auditing to Stop Reward Hacking
A new preprint proposes ARA, a dynamic two-stage framework that pairs a hacker with an auditor to catch reward hacking during RLHF and reports strong cross-domain generalization. Independent verification and compute-overhead analysis remain absent....
Read article →Simplex Diffusion Models seek to curb information collapse in discrete diffusion
SDMs lift the process to a probability simplex, carrying uncertainty across denoising steps. Unlike Dirichlet Flow Matching, which requires integrating an ordinary differential equation, SDMs use a tunable stochastic sampler....
Read article →OpenAI and SBDC Partner on Small-Business AI Training Pilot
The pilot aims to reach at least 1,000 small businesses via in-person workshops, the company said in a blog post....
Read article →NVIDIA Opens 2027–2028 Graduate Fellowship Applications With $60,000 Awards
The 26-year-old program continues NVIDIA’s investment in academic research talent, with mandatory internships preceding the fellowship year....
Read article →RAPTOR Probe Extracts Concept Vectors for LLM Steering, Preprint Claims
A new ridge-regularized logistic probe claims to produce directionally stable concept vectors at lower cost for LLM activation steering, but the work is still a preprint without external validation....
Read article →Nvidia Launches Open Agent Safety Platform, Says It Can Quarantine Rogue Agents in Milliseconds
The launch responds to what Nvidia calls “recent security incidents” in which agents circumvented software-level controls, the company said in a press release, though the company provided no independent test results for the new platform’s effectiveness....
Read article →Preprint Sets Generalization Bounds for OPTQ Quantization
OPTQ progressively quantizes weights to minimize the squared quantization error on a calibration dataset, according to the abstract of the paper arXiv:2609. 31560....
Read article →