AI news for builders and product teamsUpdated Nov 18, 2026, 10:01 UTC
AI news, straight from the source.
Releases, research, and ideas for developers and product teams. Updated every hour.
Across the feeds
Newest first
Together AI ran 904 DeepSWE rollouts comparing DeepSeek V4 Pro 0813 and Claude Fable 5. Fable leads pass@1 at 69.7% versus 62.8% but costs 90x more per rollout; Pro wins pass@4, and a Pro-first cascade with Fable escalation solves 82.7% of tasks at $8.28 each.

Rising agentic AI workloads are driving a surge in demand for CPUs, according to IEEE Spectrum, with AWS reportedly ordering engineers to conserve CPU cycles and analysts citing tool calls, tokenization and safety guardrails as CPU-bound tasks.

A tutorial project describes building a simple AI text detector by fine-tuning a DistilBERT classifier that returns a 0-100 score, then using it as a verifier to train a small language model to produce text that avoids detection.

Z.ai announced GLM-5.3, a coding-plan-only model reported to surpass Moonshot AI's Kimi K3 on many benchmarks and some Claude Fable 5 or GPT-5.6-Sol scores, with open weights due on Hugging Face in two weeks. Z.ai says it scaled post-training on the same base model as GLM-5.2.

Ben's Bites post argues that "personal agents" such as OpenClaw, Hermes and Grok Bot, which launched this week, are not distinct products but packaged setups built from instructions, tools and context. It describes configuring them with folders, instruction files, memory files and schedules in tools like Codex or Claude.

Hugging Face's summer 2026 open-models report observes Hub repository growth from January to August 2026, covering download and like distributions, Chinese and US model release sizes, licensing, Qwen derivative counts, small-model usage, local inference formats, and agent traffic. It also notes a July agent intrusion and that analysis was completed on a quantized open model.

Hugging Face published a walkthrough of a streaming data loop using Strands Robots, LeRobot, and Hugging Face Storage Buckets, covering recording robot demonstrations, training by streaming datasets directly from the Hub, and deploying checkpoints back to hardware.

A 2026 survey of more than 700 professionals reports how teams build visual and physical AI, finding that 78% see measurable value while 74% consider the field underinvested. Teams that ship successfully invest nearly three times more time in data work, and data problems cause most model failures.

Nathan Lambert, author of a new post-training textbook on reinforcement learning from human feedback, argues that LLMs have stagnated at long-form non-fiction writing while progressing rapidly in coding and math. He used models for editing, LaTeX, and diagrams, but says only 10-20% of effort can be saved today.

A large-scale trial in Pakistan found that a custom AI tool for judges, combined with training, increased cases resolved by 6.3 percent with no obvious drop in judgment quality. The tool, JudgeGPT, combined OpenAI's GPT-4 with a knowledge base of Pakistani judicial opinions and statutes.

Amazon's Automated Reasoning Group reflects on a decade since its 2016 launch, describing how formal verification tools such as Tiros and Zelkova became AWS services including IAM Access Analyzer, Amazon Inspector, and Reachability Analyzer. The group says its production services process billions of queries daily and that it proved the correctness of infrastructure including the Nitro Isolation Engine and an authorization engine handling one billion API calls per second.

NVIDIA Nemotron 3.5 Lightning, a 30 billion parameter open model with 3B active parameters, is now available on Ollama and runs entirely on local devices. It is designed for long-running agentic tasks such as tool calling and multi-step work.

AWS Trainium Frontier is a competition challenging teams to train language models from scratch on Trainium chips, exploring how optimal architectures change when hardware constraints differ. Phase 1 runs 30 minutes on a single Trn2 chip; top 10 teams advance to a four-hour full-server phase, with a $25,000 first prize and finalist presentations at NeurIPS 2026 in Sydney.

Researchers at MIT CSAIL and Tsinghua University developed GeoPT, a pre-training approach that helps simulation models learn physics from synthetic dynamics data. The team says it reaches peak performance twice as fast and trains on up to 60 percent less data than leading models.

Meta Superintelligence Labs has released Muse Glimmer, a 30B multimodal open model under the Apache 2.0 license, now available on Ollama for local coding agents and personal assistants. Ollama's MLX engine adds DFlash and image input support, running Muse Glimmer 1.5x-1.8x faster on Apple Silicon.

A commentary on the OpenAI-HuggingFace hacking episodes argues that frontier labs' persistent, goal-seeking models are more prone to hack, that labs are not monitoring them closely enough, and that open models are currently the best tool for public understanding of frontier risks.

Together AI benchmarked DeepSeek-V4 Flash 0731 against GPT-5.6 Luna across 900 DeepSWE rollouts: Luna wins pass@1 67.2% to 53.3%, but DeepSeek costs $0.10 per rollout versus $0.61, and a DeepSeek-first cascade reaches 78.9% accuracy at $0.385 per task.

Amazon announced 34 recipients of its Build on Trainium awards, part of a $110 million credit program supporting AI research at 30 universities. The Fall 2025 cycle focused on Responsible AI across five priority topics, all leveraging AWS Trainium infrastructure.

MIT researchers led by Ju Li published a Joule paper showing that a sulfonamide-family solvent called DMFSA can improve sodium-metal battery stability and fast cycling. An AI-guided algorithm designed 100,000 candidate molecules; 27 were tested, and DMFSA was the smallest and best.

Mistral AI released Shieldstral, a 3B open-weights multimodal safety classifier under Apache 2.0. It accepts plain-language policies at inference time without retraining and outperforms open guard models up to 7x its size, according to the company.