Loading activity…
@nrehiew_
eng primarily, ml mostly, research previously
AI
The reported MATH-500 score rose from 7.9% to 48.1% with the same sampling settings.
AI
Core Automation describes two sites with different GPUs and parallelism schemes connected by a bidirectional 25 Gbps link.
AI
Cognition says its 132 tasks were made by repository maintainers; its leaderboard puts Claude Opus 5.5 first at 45.5%.
AI
OpenAI says agent-led tests exposed VM escapes and limited internet routes, leading to patches and stronger monitoring.
AI
Kimi claims a 1-million-token context window for Pro and two to three times Pro's speed for Highspeed.
XAI
A user calculates the range assuming 97% cached input tokens and GPU rental costs of $3–$6 per hour.
AI
DeepSeek's README offers offline inference for DeepSeek-V4.1-Flash but lists additional systems needed for a complete service.
AI
PrimeIntellect says IPO is tailored for stability in asynchronous reinforcement learning.
AI
DeepSeek says the model is live on its website, app and API, and optimized for DeepSeekHarness and Tencent CodeBuddy.
AI
A user shared notes on Score Centering and predicted it would become part of the standard reinforcement-learning recipe.
AI
A user argues that trainer–sampler mismatch creates a drift term that can destabilize reinforcement-learning training.
AI
A user argues that publishing such blogs shows no respect for people interested in the company.
AI
A user flags HBM4 and the absence of super-long-context testing as caveats to Jalapeño–Blackwell comparisons.
AI
A ProximalHQ team member says they dynamically adjust the inference-trainer ratio to maintain training throughput and reduce staleness.
XAI
CoreWeave says RL Rollouts updates live deployments during reinforcement-learning training without touching in-flight requests.
AI
OpenAI says it drew on an independent mathematics advisory group’s advice about how to release the results.
AI
The newcomer says they see open problems in data and post-training and invites people interested in large-scale infrastructure to reach out.
AI
Google announces a 1 million-token output limit, up from 64,000, for its new model. Initial access goes to trusted cyber defenders through the Fairwind Program.
AI
A user praises a team and congratulates Calvin Chen.
AI
Xiaomi says Pro performs on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks, with stronger coding, computer use and 3D reasoning capabilities.
