Researchers propose Adaptive Reward Routing, a method that dynamically adjusts update locations and reward coordination during the training of joint audio-video diffusion models. ·
Researchers find that token-level KL direction shapes performance in Llama3 and Qwen2.5 distillation, while rollout policy plays a less central role. ·
Researchers propose Adaptive Reward Routing, a method that dynamically adjusts update locations and reward coordination during the training of joint audio-video diffusion models. ·
Researchers find that token-level KL direction shapes performance in Llama3 and Qwen2.5 distillation, while rollout policy plays a less central role. ·
DeepSeek Harness is a desktop application for macOS and Windows that processes documents, spreadsheets, and code while scheduling tasks through composable plugins. ·
Researchers propose Adaptive Reward Routing, a method that dynamically adjusts update locations and reward coordination during the training of joint audio-video diffusion models. ·
Researchers propose Adaptive Reward Routing, a method that dynamically adjusts update locations and reward coordination during the training of joint audio-video diffusion models. ·
Researchers find that token-level KL direction shapes performance in Llama3 and Qwen2.5 distillation, while rollout policy plays a less central role. ·
Researchers introduce OneStreamer, a streaming video model that records evidence and responds proactively using hierarchical caption memory and state transition learning. ·
HF Daily Papers21h ago▲arxiv.org
X voices
5 stories
X posts other sources are linking to, then named voices on Bluesky and Mastodon
OpenAI releases Codex version 0.160.0, which adds keyboard-accessible browsing for older tasks and middle-click pasting in fullscreen mode on Linux terminals. ·
OpenAI releases Codex version 0.160.0, which adds keyboard-accessible browsing for older tasks and middle-click pasting in fullscreen mode on Linux terminals. ·
BootLoops 1.0 provides certified computational tools for exact and high-precision physics and quantitative science, designed to be driven by LLM agents under an MIT license. ·
Unsloth releases a desktop update adding a command palette, shareable settings, and faster Laya decisions. It also reduces peak memory usage by 45% for NVFP4 LoRA training. ·
GitHub Releases31h ago▲github.com
Infrastructure
5 stories
Data centers, power, chips and what tokens cost
Data-center and energy stories per day
Table
Sat 26
1
Sun 27
1
Mon 28
20
Tue 29
33
Wed 30
34
Thu 01
26
Fri 02
32
Stories from DCD, Data Center Knowledge, EIA and Hacker News, by publication day.
Median output price of models added in the last 30 days, $ per million tokens
One point per OpenRouter reading, last 8 days. Data by OpenRouter, CC BY 4.0.
Output price, $ per million tokens (newest models)
Table
pareto-26.10-preview
3.2
gpt-6.1-sol-pro
10
gpt-6.1-sol
10
claude-sonnet-5.5
10
claude-sonnet-5.5:batch
5
median, all paid models
2
437 paid models listed; median of the 52 added in 30 days: $5.00. Rankings data by OpenRouter, CC BY 4.0.
Russian strikes take another data center in Kyiv offline, specifically the Dline.ua facility, following attacks on numerous other data centers in the city. ·