AI Coding Live News

Theo (@theo), CEO of T3 Chat and T3 Code and a well-known developer YouTuber, posted on November 8 that LLMs are "still SO bad at building tools for agents": every time he asks Opus to set up an MCP authentication path, it "makes horrible assumptions and builds something that straight up doesn't work." About two hours later he added that he is writing an "MCP the right way" guide for his own agents; when he had them draft a first version, they proposed rendering the MCP URL as a clickable link for users, and he warned people not to trust "ANYTHING" agents say about MCP, connection strategies and agent "best practices." At writing time the first post had about 880 likes, 136 replies and 65,000 views.

Zoom

On November 8 (morning of November 9 in Singapore), independent evaluator Artificial Analysis released version 2 of its Decision Index for decision models. The headline metric is now a Decision Score that rewards both correct answers and well-calibrated confidence, the private-data share rises from 25% in v1 to 45%, and two new evaluations are added: AA-Paperclip with 1,500 private documents and AA-Symmetry for answer consistency. Among decision models, Nace AI's Drex Max v1.1 and Fastino Labs' GLiDE v1.2 both score 73, with Drex Max narrowly ahead, and Base Labs' Mammoth 27B scores 72. General LLMs score higher, with Claude Haiku 5.5 (Max) at 77 and GPT-6 Luna (Max) at 76, but cost $6,240 and $1,530 per 1M tasks versus $102 for Drex Max.

Zoom

CellType, a YC W26 startup, released BioSimBench 1.0 on November 4. It uses real lab experiments to test whether language models can predict which genes go up or down after a cell is hit by a drug, a gene knockdown or an immune signal. Among 12 frontier models, Gemini 4 Argon ranks first with an overall score of 0.184, ahead of Claude Sonnet 5.5 (0.160) and GPT-6 Astra (0.159), while two lab replicates agree at 0.462. A 27B model CellType trained in its BioSimGym RL environments beats the best frontier model on all four prediction tracks.

Zoom

On October 29, Randall Balestriero and Yann LeCun posted "Proving and Improving the Platonic Representation Hypothesis with JEPAs" (arXiv 2610.39302), affiliated with Brown University, NYU and AMI Labs. The paper proves that JEPAs trained separately on different modalities learn the same representation up to one rotation whenever their positive pairs are two observations of one scene ("good views"), and argues that next-token prediction is a JEPA with "bad views", which is why language models only align with vision through scale. The authors report that good views need 42x fewer training samples than next-token prediction to reach the same alignment, and that a 109M text JEPA scores 0.184 alignment with DINOv2 versus 0.154 for Qwen2.5-72B (author-reported). Code and a live demo are public; Balestriero's announcement post has about 1,054 likes and 72K views.

Zoom

On 2026-10-23 (16:53 UTC) the official @blocks account unveiled Dessa, and Jack Dorsey called it the most authentic full-duplex speech model ever built. The official page block.xyz/dessa says Dessa is a true speech-to-speech model trained only on recordings of human conversation, with no text and no pretrained models. It is full-duplex, so it listens and talks at the same time, cuts in and backchannels naturally, runs faster than real time, and runs locally on a MacBook Pro M4 Max or better. Block says it was trained on a single NVIDIA GB200 NVL72 and learned language from hundreds of times fewer words than LLMs. There is no public API or weights yet; a small group of AI safety researchers is getting early access.

Zoom

Artificial Analysis launched AA-PersonalityBench (@ArtificialAnlys), an innovative behavioral benchmark quantifying psychological traits in frontier LLMs, with initial rankings focusing on Confidence (uncertainty calibration) and Sycophancy (tendency to agree with user misconceptions). The data revealed that Google's Gemini 3.8 Flash (100) and Gemini 4 Argon (97) displayed the highest assertiveness while being the least likely to admit uncertainty. Conversely, on sycophancy tests, Llama 4 Maverick scored 97 (frequently validating flawed user prompts), whereas Claude Fable 5.1 scored a near-flawless 4, refusing to flatter and rigorously holding factual ground.

Zoom

The vLLM project team and Ant Group officially open-sourced Humming (humming-kernels), a specialized high-performance kernel library designed for serving quantized MoE and dense LLMs at massive scale. Humming natively supports WnA16, W8A8, W4A8, and W4A4 across INT1–INT8 and FP4–FP8 formats, spanning NVIDIA architectures from Turing through Blackwell as well as T-Head Zhenwu 810E silicon. In production benchmarks, Humming kernels outpace cuBLAS by 3–4x at small batch sizes and surpass Marlin by 2x on H200 clusters. Serving DeepSeek-V4.1-Flash across 8x H20 GPUs, Humming drives a +43.3% throughput gain under 64 concurrent requests while reducing time-to-first-token (TTFT) by 32%, pre-integrated directly into vLLM v0.29.0+.

Zoom

While supervised fine-tuning (SFT) distills agentic tool-use capabilities from frontier LLMs into small language models (SLMs, 0.6B to 8B parameters), student agents frequently suffer from 'progress stagnation'—collapsing into repetitive, unproductive tool-calling loops as interaction trajectories extend. In arXiv:2610.19439, researchers introduced Progress-Aware Distillation (ProD). ProD defines a rigorous stagnation scoring formulation that dynamically identifies and penalizes stagnant behavior during student trajectory rollouts via an adaptive SimPO-inspired margin. Experiments across 8 diverse SLMs prove ProD effectively eradicates repetitive loops while boosting task completion efficiency and success rates.

Zoom

Researchers introduced HERA in arXiv:2610.06563, addressing a critical reliability gap in autonomous LLM agents: the inability to detect task infeasibility and reliably abstain from blind, destructive tool execution. Rather than tuning harnesses against static benchmarks that quickly saturate, HERA establishes a harness-environment co-evolution framework that mutates execution environments to expose emergent failure modes while adapting the scaffold. Evaluated across 19 frontier LLMs using the newly released hera-bench, HERA elevates agentic abstention success from 61.7% to 83.3% and paired task success from 48.3% to 70.0%.

Zoom

Researchers from Stanford University and UC Berkeley introduced Librarian Agents (arXiv:2610.22273). To overcome the bottleneck of reactive RAG methods struggling with complex cross-document reasoning at query time, the framework uses reinforcement learning to train LLMs to perform anticipatory directed curation on messy long-context workspaces into structured synthetic filesystems. On BrowseComp+, it hits 55.86% accuracy while reducing navigation token consumption by 33.8% to 51%.

Zoom

Multimodal LLMs excel at holistic image captioning but struggle with pixel-level spatial grounding due to lack of spatial supervision. Xiaomi AI Lab unveiled Xiaomi-Lens-0 (arXiv:2610.23279), a 4B MLLM based on Qwen3.5 that natively predicts pixel masks. A lightweight decoder integrates patch embeddings, visual latents, and mask-token activations, driving RefCOCO comprehension (REC [email protected]: 94.6% / 92.1% / 92.5%) and segmentation (RES cIoU: 83.4% / 79.9% / 81.1%) past Florence-2 and LISA-7B while preserving a 69.7% MME-RW score.

Zoom

Continuous diffusion language models (CDLMs) currently transplant architectures from causal LLMs or vision DiTs that are not optimized for parallel diffusion. Researchers from Xiaohongshu (Dots Studio), Peking University, CUHK Shenzhen, and UCLA introduced LangDiT (arXiv:2610.22330). By separating feature normalization from noise modulation, detaching shortcut gradients, and introducing confidence-gated self-conditioning, LangDiT reduces generation perplexity (GenPPL) to 21.72 on LangFlow and 11.33 on Plaid across 1024-token OpenWebText rollouts.

Zoom