Daily bookmarks: relevant news, facts and discoveries.
Articles 1216
Engineering ·
A recent arXiv paper argues that autonomous science depends on physical validation capacity and proposes markets and institutions to allocate that scarce work.
Engineering ·
Intel presents the GPU D-Series (formerly Crescent Island), with up to 480GB of LPDDR5X at 1.5TB/s, already sampling. Radar briefing with the specifications stated in the source.
Engineering ·
Open-source project and paper describing μLattice, a memory design that grows with input without quadratic cost. The author claims fewer operations than attention and perfect recall of thousands of associations.
Engineering ·
Radar: the QwenGyre framework, described by the Qwen team at Alibaba, reallocates GPUs between rollout and training and groups branched agent runs into trajectory trees, with a reported speedup of up to 1.85x over Colocate and about 44% less training time.
Engineering ·
A more elaborate workflow does not always improve outcomes. Compare components by task quality, consumption and data exposure.
Engineering ·
An arXiv paper proposes HERMES, a code-agent harness in which each repository component has a resident LLM. The authors report gains in migration and on four software engineering benchmarks against comparable baseline harnesses.
Engineering ·
According to Artificial Analysis, a commercial voice service ranked first in the Controlled Voice TTS Arena, with an Elo of 1,201 across eight cloned voices in US and UK English.
Engineering ·
A post presents an open-source 1.7B text-to-speech model for voice cloning, with controls for emotion, age, tone, accent and multilingual support. Its realism and latency claims remain unverified.
Engineering ·
Cactus-Compute published Whistle, a 16.9 MB speech-to-text model for on-device use. According to the author, it rivals Whisper base for that kind of use.
Engineering ·
Stanford article on DeLM, which replaces a central coordinator with shared context and a task queue for long-horizon code agents. According to the author, it is up to 2.49× faster and more accurate than the cited baselines.
Engineering ·
A research item proposes LayerRoPE, which encodes depth in normalization weights to handle growth of hidden-state norm across layers in very deep Transformers.
Engineering ·
According to the report, Samsung open-sourced LittleBit, a method that uses latent factorization to shrink weights of 13 billion parameter models to under 1 GB, with sub-1-bit levels and XOR in place of multiplication.
Engineering ·
Meta presents RoboJEPA, a JEPA-based latent world model scaled from 22M to 8B parameters and trained on data from 12 robotic embodiments. The paper studies whether latent prediction error predicts later planning performance.
Engineering ·
A social media post describes RoboJEPA, an 8B-parameter JEPA world model trained on 15,022 hours of robotic video, and reports scaling-law adjustments that predict results for its 4B and 8B variants.
Engineering ·
Site comparing accuracy and speed of speech-to-text models supported by transcribe.cpp, measured per language on FLEURS, with data from tests on a 4750U laptop.
Engineering ·
A NeurIPS 2026 paper cloned the voices of 86 non-native English speakers. In 4,000 human ratings, the clones were judged warmer, more authoritative and more human.
Engineering ·
Radar: a paper led by Yann LeCun proposes H-JEPA, which stacks JEPA world models across multiple time scales for long-horizon visual planning.
Engineering ·
Radar: LiFT trains an image generator with flow matching and a shared transformer core, where each loop progressively refines the initial velocity estimate. According to the author, a model trained with 2 loops improves with up to 16 loops at inference.
Engineering ·
A post reports a single-voice, single-language text-to-speech model of about 9MB of weights, distilled from Kokoro-82M, which has 8M parameters. The author claims it runs anywhere and is very fast, without independent verification.
Engineering ·
A post reports removing up to 10 layers from Qwen3-8B-Base with LaCo. With 6 layers removed, the model is 16.7% smaller, keeping 78% to 94% of reasoning scores with a perplexity ratio of 1.68x, according to the author.
Looking for something else?DocumentationContact
