Forward-Only Training Boosts LLM Fine-tuning 3x, Cuts Memory 40%
Forward-Only Training Boosts LLM Fine-tuning 3x, Cuts Memory 40%
A new arXiv paper introduces Forward-Pass-Only (FPO) domain adaptation, a method that trains LLMs without backpropagation through the model body. It achieves 2.7-3.2x the throughput of standard fine-tuning while using roughly 40% less peak memory, with off-domain benchmarks remaining stable. This could enable efficient model adaptation for resource-constrained teams.
BCMT: Blockwise Memory Architecture Beats Quadratic Attention
A new arXiv paper introduces BCMT (Blockwise Causal Memory Transformer), a novel architecture designed for long sequences. Standard self-attention has quadratic complexity with sequence length; BCMT replaces dense attention with blockwise causal memory, significantly reducing computational cost. This offers a more efficient architecture choice for LLMs handling ultra-long contexts.
DumpsterCluster: Serving LLaMA-70B on $60 Retired GPUs
A new arXiv paper explores building a DumpsterCluster from retired datacenter GPUs to serve LLaMA-70B inference. As AI datacenters retire functional hardware, these accelerators flood secondary markets. The study shows that with proper distributed strategies, low-cost recycled GPUs can still power modern LLM inference, dramatically reducing serving costs.
Jais 2: New Open Arabic-Centric LLM Family Released
MBZUAI, Cerebras, and Inception have jointly released Jais 2, a family of Arabic-centric large language models with strong performance on Arabic and culturally grounded benchmarks. The open models aim to advance Arabic-centric language modeling. Paper: arXiv:2608.13580v1.
SKILL: Self-Correcting LLM Agent for Logic Optimization
A new arXiv paper introduces SKILL, a framework enabling LLM agents to perform iterative logic optimization through knowledge-guided self-correction. Traditional expert-designed flows lack adaptability, while RL methods struggle with large search spaces and sparse rewards. SKILL leverages LLM reasoning to address these challenges, offering a fresh approach to logic synthesis in chip design automation.
Cochain-Frame Equivariance: A New Way to Learn Discrete Physics
A new arXiv paper presents a learning method based on Cochain-Frame equivariance for physical fields on meshes. These fields require separating topology from geometry: conservation laws are topological and exact, while geometry and material responses must be learned. The approach improves accuracy and generalization of neural surrogates in physical field prediction.
Mojo Programming Language Goes Open Source Under Apache 2
Mojo has fulfilled its promise from May 2023 by open-sourcing its compiler and toolchain under the Apache 2 license, following its 1.0 release. Developers can now freely use, modify, and distribute this language designed to combine Python’s usability with C-level performance.
Chat SDK Adds Instagram Adapter for DM Bots
Vercel’s Chat SDK now includes an Instagram adapter, letting developers build bots for Instagram that can send and receive DMs and media, render cards as quick replies, link buttons, show typing indicators, react to messages, and handle story replies via Meta’s Instagram Messaging API.
Vercel Offers $1M Bounty to Hack Its Sandbox Isolation
Vercel has launched a $1 million hacker challenge inviting security researchers to attempt breaking into its Sandbox product. Sandbox uses microVM technology with a dedicated guest kernel per workload, isolated from the host and other workloads. Given recent research highlighting risks in running untrusted code, Vercel aims to validate and harden its isolation through crowdsourced testing.
Asana Uses Codex to Finish 5 Years of Work in 2 Weeks
Asana used OpenAI Codex to replace an outdated testing system in just two weeks, completing work expected to take five years at a cost of approximately $12K. This is another real-world case of AI coding tools tackling large-scale legacy system modernization, demonstrating the efficiency of code agents in production engineering tasks.
OpenAI Launches Initiative for Democratic AI Oversight in National Security
OpenAI has launched an initiative to strengthen democratic oversight of AI in national security. The program supports government institutions with tools, training, and expertise to ensure AI deployment in sensitive areas aligns with democratic values and accountability.
OpenAI Tightens Security to Pace Frontier Model Development
OpenAI is strengthening monitoring, alignment, and security for frontier AI models, with new safeguards now guiding the pace of development. As capabilities grow, the company is adopting a more cautious approach to balance progress with risk, ensuring each stage of expansion is thoroughly evaluated. This responds to growing industry attention on AI safety governance.
OpenAI Launches ChatGPT for Teens with Parental Controls
OpenAI has officially launched ChatGPT for Teens, providing AI-assisted learning for adolescent users. The version comes with stronger built-in protections, healthy-use features, and additional controls for parents, enabling teens to learn and think critically in a safe environment. This marks OpenAI’s further expansion into educational scenarios.
Glean CEO: Model Routing Key to Controlling AI Costs
Glean CEO Arvind Jain explains in a Latent Space post why model routing helps organizations control AI costs. He highlights that large-scale human feedback loops can continuously improve routing systems, reducing LLM API expenses while maintaining output quality.
Would Your Personal AI Agent Benefit From Desktop History?
Ben’s Bites explores the concept of giving AI a complete history of your desktop activity. This would allow personal agents to better understand work habits and context, offering more personalized assistance. The piece discusses possibilities and implications.
Vercel for Platforms Now Deploys Directly From Users' GitHub Repos
Vercel for Platforms now enables teams to create deployments directly from users’ GitHub repositories without requiring them to install the Vercel GitHub App. Pass a gitAccessToken alongside gitSource when creating a deployment, and Vercel uses the token to fetch and build.
Claude Code v2.1.235 Adds Spellcheck, Fixes Cache Invalidation
Claude Code v2.1.235 adds an optional spellcheck setting that underlines misspelled words in prompts, using aspell, hunspell, or ispell. It also fixes whole-prompt-cache invalidation caused by language server disconnects mid-session, plus a nested markdown list rendering bug.
OpenAI Codex Ships rust-v0.149.0-alpha.1 and Other Updates
OpenAI Codex has released multiple version updates: rust-v0.149.0-alpha.1, 0.148.0, 0.148.0-alpha.23, 0.148.0-alpha.22, and python-v0.147.0. The rust-v0.149.0-alpha.1 is the latest release.
OpenAI Partners With CodeAI to Build AI Literacy in Students
OpenAI has partnered with CodeAI to help students build AI literacy, think critically about AI, and develop skills to use and shape it responsibly. The collaboration prepares the next generation for an AI-pervasive society.