AI Coding Live News

Over the past day, X has been debating whether GPT-6 Astra can work directly as a robot policy. Former OpenAI researcher Will Depue (whose bio lists Sora 1 and 2 and o3/4o post-training) called Astra beating specialized robotics models "deeply surprising", then described it as "the final death of specialized ML". Investor Freda Duan posted that "Astra IS robotics' GPT-3 moment", drawing about 84K views. Third-party evaluations point both ways: in a RoboDojo paper, Astra averages 22.48% success and ranks above all 40 public policies, while an EBench report puts it at 46.7%, third of nine models. This is a discussion of opinions, not an official OpenAI conclusion.

Zoom

On Oct 26 Mirendil unveiled Fern, its first model, pretrained from scratch on its own Mirendil-Base (0.1-0.5T params) with a context window of up to ~10M tokens, about 0.5M lines of code or days of agent traces. Fern 0.1 got only a short round of post-training and targets AI R&D work such as auditing training data, debugging runs from logs and reading agent traces end to end. Vendor-reported: 73.4% on SWE-Bench Verified and 100% single-needle RULER retrieval at 10M, while multi-hop QA drops to 45% at 10M. Access is early-access only; no price or weights yet.

Zoom

On Oct 19 (UTC) fal released FLUX 3 Max, a video model built by fal Research with Black Forest Labs through post-training, reinforcement learning and inference optimization on top of the FLUX 3 foundation model. It handles text-to-video, image-to-video and keyframes, up to 20 seconds, at 720p or 1080p, with audio on by default. On Artificial Analysis's preliminary image-to-video arena it scores 1,139 Elo, level with Seedance 2.5 and within the confidence interval of leader MiniMax H3 (1,141), and fal says a 10-second clip takes about 12.45 seconds. A 75% launch discount runs until Nov 7, putting 720p at $0.0425 per second.

Zoom

Three new frameworks advance web and multimodal agent post-training: Salesforce AI Research's CLIFT (arXiv:2610.06829) certifies URL-conditional self-verification questions for dense RL rewards and judge-free Conformal Trajectory Selection (CTS), propelling Gemma-4 31B to a 74.6% SOTA on WebArena Infinity (beating Gemini 3 Flash at 70.1%) and lifting GPT-5.5 to 53.7% on VisualWebArena. Meanwhile, MMSearch-OPD (arXiv:2610.22240) boosts 9B multimodal deep search by +15.3 points to 60.2%, and Northeastern University's SCOPED-8B (arXiv:2610.22217) uses skill-decoupled on-policy expert distillation to reach 73.13% across 8 benchmarks—surpassing Qwen3-VL-32B (72.01%) and Claude Opus 4.6 (72.86%).

Zoom

Ollama launched US- and EU-hosted Ollama Cloud with strict zero data retention and zero training, adding immediate support for DeepSeek-V4.1-Flash-1017 via one-line launch commands for Claude Code and OpenCode alongside Ollama v0.40.6. Meanwhile, SemiAnalysis revealed that GLM-5.3 was built entirely by post-training the previous GLM-5.2 base model without pretraining from scratch, achieving a 'night and day difference' in cybersecurity and agentic capabilities.

Zoom

Alibaba researchers introduce Cowork-Gym (arXiv:2610.17670), synthesizing 64,900 interactive human-agent coworking environments paired with a stateful user simulator that dynamically edits the shared workspace; post-training a 27B Qwen model yields #1 scores on PawBench and ClawEval, outperforming 100B+ parameter models. Concurrently, Harvard and MIT researchers open-source Market-Based Post-Training (MPT, arXiv:2610.22322), enabling decentralized agent swarms to jointly learn action policies and bidding-based self-orchestration without a central coordinator.

Zoom

To test whether frontier agents can autonomously conduct end-to-end post-training research, researchers introduce Athanor, an agent-native system for autonomous post-training (AutoPostTrain) on real GPUs. In a 104-hour campaign without human advice or external teacher models, four Claude Opus 5.5 agents—each operating an 8-GPU node—improved raw Qwen3-4B-Base from 15.9 to 67.3 on the dev suite and from 9.2 to 48.2 on an unseen held-out suite, surpassing the official Qwen3-4B-Instruct-2507 (66.5 and 47.9).

Zoom

On 2026-10-15 Microsoft AI published a technical post and PDF saying it has finished pre-training its in-house base model MAI-Base-2: MoE in every layer (1 shared + 7-of-512 routed experts), 80B active / 2.8T total, 102 layers, native encoder-free vision, a 50T-token mixture and a March 2026 knowledge cutoff. Per MAI, it beats every open-weight model tested on held-out BPB and its recipe is 4.2–12.3x more pre-training-FLOPs-efficient than DeepSeek-V4-Pro-Base. It is still in post-training: no API, no weights and no release date yet; Microsoft plans to license weights to trusted partners via Foundry.

Zoom

San Francisco lab Paper Instruments launched Ultramarine 1 Flash on 2026-10-14 16:02 UTC: GLM-5.3 Flash post-trained with large-scale non-verifiable RL (per-task rubric rewards) for finance, legal, consulting and operations work. Its report claims +4.1 points on average over the base across six knowledge-work benchmarks with 6.5% fewer output tokens, and 59.3 at $0.13/task on its internal PI Knowledge Work suite vs Claude Fable 5 at 57.8 / $8.19. An OpenAI-compatible API (ultramarine-1-flash) is live; the CEO says weights follow in about two weeks and post-training cost ~$300k. The headline cost comparison rests on an internal benchmark and has not been independently reproduced.

Zoom

While LLM agents are rapidly deployed into sandbox workspaces with direct shell and filesystem access, post-training methods remain predominantly one-shot and bound by external teacher models. ClawLoop introduces a self-improving, harness-native training loop that operates without external teachers. Through a self-sustaining cycle of task synthesis with difficulty growth, SFT/RL training, and automated behavioral diagnosis, ClawLoop demonstrates consistent empirical gains across SkillsBench, Claw-Eval-Live, and ClawGym-Bench on Qwen3.6-27B, while successfully transferring data to 4B and 9B models.

Zoom

Reinforcement learning (RL) post-training is foundational to scaling reasoning capabilities in large language models. However, its memory footprint remains a prohibitive bottleneck: storing full-sized gradient buffers alongside multi-state Adam optimizer moments frequently triggers out-of-memory (OOM) failures when attempting to train large models (such as 27B parameters) on standard compute hardware. In newly released research (arXiv:2610.06647), researchers introduce LoGRA, an efficient RL post-training framework available within the open-source Molt library (github.com/skzhang1/labs-molt). LoGRA compresses training representations by retaining useful learning signals inside compact low-rank gradient sketches accumulated directly during backpropagation, facilitating both memory-efficient parameter updates and streamlined policy synchronization. To counteract instability from gradient compression, LoGRA introduces a Predicted-KL step control mechanism that forecasts policy change magnitudes prior to parameter updates, dynamically attenuating disruptive updates. Evaluated across reasoning benchmarks, LoGRA reduces average training memory by up to 45.7% with zero degradation in task performance, successfully sustaining stable RL training of a 27B model for over 1,100 steps on a single eight-GPU node where standard dense Adam crashes with OOM.

Zoom