AI Coding Live News — Real-Time 𝕏/Twitter Updates & Breakthroughs

Gökdeniz GülmezGoekdeniz-Guelmez·Collected 36d ago

👀96

🔥78

MLX-LM-LoRA v5.8.6: Apple Silicon fine-tuning library jumps from 3.1.3, adds DSLA preference alignment and critic-free KLPO RL

Gökdeniz Gülmez shipped MLX-LM-LoRA v5.8.6 on Oct 5 (GitHub Release and PyPI), the first release after 3.1.3. It adds DSLA with DPO/ORPO/CPO objectives and critic-free KLPO with token/sequence routes and several KL estimators, improves GRPO stability and microbatching for Online DPO, XPO, RLHF REINFORCE and PPO, and adds fast VJP for gated-delta layers. Minimums rise to mlx>=0.32.3, mlx_lm>=0.32.0 and Python>=3.11; license standardized on Apache 2.0.

Key Takeaways

Jump from v3.1.3 (Sep 16) straight to v5.8.6, merging nine PRs (#64, #94, #96–#102).
Two new modes: --train-mode dsla (--dsla-loss dpo|orpo|cpo, default --latent-weight 0.1) and --train-mode klpo (--klpo-route token|sequence, estimators mc/binary/topk/full).
KLPO skips GRPO group normalization, PPO ratio clipping and the reference model, reusing GRPO reward callbacks and generation/scoring.
Raised minimums: mlx>=0.32.3, mlx_lm>=0.32.0, Python >=3.11.
QAT now covers DSLA; new GitHub Pages site with a Python API reference.

Read full story

Linked Models & Agents:

🌐 apple

#MLX#MLX-LM-LoRA#Apple Silicon#Fine-tuning#DPO#RLHF#KLPO#DSLA

Amazon AI@awscloud·Collected 45d ago

👀169

🔥94

Amazon Open-Sources Rufus-Air: 8-Stage Post-Training RL Recipe on GLM-4.5-Air MoE Surpassing Official Checkpoints

Amazon's AI Research division has open-sourced 'Rufus-Air: An Open LLM Post-Training Recipe' (arXiv: 2609.29421), an entirely reproducible eight-stage post-training pipeline built atop GLM-4.5-Air-Base (a 106B total / 12B active parameter Mixture-of-Experts model). Dispensing with proprietary teacher distillation and private human annotation, Rufus-Air cascades through eight distinct stages—SFT, Reasoning RL, Coding RL, Instruction-Following RL, General Agent, Coding Agent, Search Agent, and RLHF—leveraging solely public data and open tooling to beat the official GLM-4.5-Air checkpoint across standard benchmarks.

Key Takeaways

Delivers the first fully reproducible 8-stage post-training recipe for 100B+ MoE architectures (106B-A12B GLM-4.5-Air).
Relies entirely on publicly available datasets and verifiable programmatic reward models without proprietary teacher distillation.
Validates the 'Reward Reliability Ordering' heuristic, sequencing stages from deterministic unit tests to soft LLM judge signals.
Outperforms the official GLM-4.5-Air checkpoint across core evaluation suites, registering a 21.8% boost on multi-turn coding agent tasks.
Full stage ordering configs, reward architectures, and distributed training recipes published openly on arXiv and Hugging Face.

Read full story

Linked Models & Agents:

⚡ GLM 4.5 Air⚡ Rufus Air🤖 coding-agent 🤖 search-agent 🌐 Amazon AWS (Nova)🌐 Zhipu AI (智谱 AI)

#Amazon#Post-Training#GLM-4.5-Air#MoE#Reinforcement Learning#Open Source#Hugging Face

Google DeepMind@GoogleDeepMind·Collected 47d ago

👀219

🔥95

Google DeepMind Enters Post-Training Phase for Gemini 4: Unified Embodied Control & Long-Horizon Reasoning

Google DeepMind (@GoogleDeepMind) confirmed that its next-generation frontier model, Gemini 4, has officially transitioned into full post-training and RLHF alignment. Built on an omnimodal world-model foundation, Gemini 4 natively unifies real-time sensory perception, robotic action tokens, and million-token reasoning for complex multi-hour agent workflows.

Key Takeaways

Gemini 4 base pre-training is finalized, moving into large-scale reinforcement learning and automated red teaming.
Directly unifies audio-visual inputs with physical action tokens, generating joint-level motor trajectories natively.
Extended context architecture supports 5,000,000 tokens with near-perfect needle-in-a-haystack recall across modal streams.
Early Explorer access will roll out through Google AI Studio for enterprise and academic partners ahead of general availability.

Read full story

Linked Models & Agents:

⚡ Gemini 4⚡ Gemini 1.5 Pro 🤖 gemini-agent 🌐 Google DeepMind

#Google DeepMind#Gemini 4#World Models#Embodied AI#TPU v6e#Research

Diogo Almeida@CompleteSkeptic·Collected 50d ago

👀62

🔥95

View on 𝕏 ↗

TypeSafe co-founder shares coding-agent design notes: routing, compaction, subagents, skills, and MCP

@CompleteSkeptic (Diogo Almeida, TypeSafe; RLHF background) published public design notes on TypeSafe × coding agents (Google Doc), covering routing, context compaction, subagents, skills, and MCP tradeoffs, and invited the community to experiment. The post drew ~930 likes / ~56k views the same day.

Key Takeaways

Public Google Doc on coding-agent routing, compaction, subagents, skills, and MCP.
Authored by a TypeSafe principal; explicitly invites community experimentation.
Same-day high engagement (~930 likes / ~56k views) as a design reference.

Read full story

Linked Models & Agents:

#TypeSafe#coding agents#MCP#Skills#Subagents#design notes