Topic: run

Inkleing 2 Small was trained in just 68 hours on 12T tokens, outperforming Muse Spark in a recent study by METR, highlighting rapid advancements in AI training efficiency. This achievement underscores the ongoing competition and progress in developing more capable AI models faster.

A new software called AegisKernel has been developed as a zero-trust POSIX runtime specifically for AI agents, executing with extremely low latency under 0.016 milliseconds. This innovation is significant because it enhances the security and efficiency of AI systems by limiting trust to only necessary components, crucial for applications requiring rapid response times and high security.

adikshatriyag-crypto.github.io↗

Zyphra released open-source FP8 and BF16 prefill attention kernels for AMD's MI355X, achieving up to 1.76x faster performance with significantly lower error rates compared to AMD’s AITER, highlighting potential advancements in GPU efficiency for AI models.

Cuban claims Facebook knowingly ran AI-generated deepfake scam ads, potentially raking in $16 billion in fraudulent revenue. This matter is significant as it highlights serious concerns over social media platforms' vulnerability to sophisticated scams and the scale of potential financial harm.

A recent study proposes pruning redundant "attention heads" in large language models can reduce their tendency to memorize and reproduce exact sequences from training data, addressing privacy and intellectual property concerns.

arxiv.org↗

The Hacker News community discussed various hardware setups for running large language models locally, highlighting the balance between performance and cost. This conversation is crucial as it helps users decide on compatible hardware, optimizing their local AI model experiences.

news.ycombinator.com↗

AI tools allow writing software via natural language, but there's often a discrepancy between intended functionality and actual outcomes, especially critical in smart contracts used in Web3 applications where precise code execution is essential. This gap highlights the need for better alignment between human intent and automated code generation to prevent errors that could lead to financial loss or security vulnerabilities.

arxiv.org↗

Factory released Droid 1.0, a model-independent coding agent that simplifies programmatic tool calling; it scored 65.2% on Terminal-Bench 4 runs with Claude Opus 5.5 at a cost of $4.73 per passed task, marking a step towards more accessible and efficient coding assistance.

marktechpost.com↗

Block's AI lab introduced Dessa, a full-duplex speech model trained solely on human conversation audio and designed to run locally on a MacBook Pro M4 Max. The release highlights advancements in privacy-focused AI models that process and generate speech directly from local devices without cloud dependency.

Ollama released version 0.40.13 of its LFM3 models, enabling text-to-speech functionality on both Linux and Windows in addition to existing platforms, expanding accessibility for users across different operating systems. This update is significant as it broadens the user base and compatibility of Liquid AI's advanced language models.