Latent Space
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Anthropic Applies Lessons From Unix and Linux to Coordinating Claude Agents
At AI Engineer New York, Anthropic explained how dynamic workflows help Claude agents tackle large projects without overloading developers, drawing on lessons from Unix and Linux.
Standard Bots Trains Industrial Robots on Demonstrations and Corrections from Factory Floors
Standard Bots is building a robot stack: pretrained models learn factory tasks from demonstrations, then improve through corrections from real deployments.
Google DeepMind and Biohub Explain Why AlphaFold Didn't Close the Protein Folding Problem
Pushmeet Kohli of DeepMind and Sal Candido of Biohub break down what AlphaFold left unsolved after the Nobel Prize and 200+ million predicted structures.
Ontologies Are Back: How AI Agents and Neo4j Are Reviving the Semantic Web
AI engineers are rediscovering ontologies, structured knowledge graphs, to keep probabilistic LLM agents inside deterministic logical boundaries.
Thinky Releases Open Inkling 975B-A41B Model Under Apache 2.0 License
Thinky has released Inkling, a 975B-parameter multimodal MoE model with open weights under Apache 2.0, plus a compact 276B version, Inkling-Small.
Lila Sciences: Science as Data for AI — an Autonomous Lab Backed by $550M
Lila Sciences is building autonomous labs where AI runs experiments through robots, betting that science is the last inexhaustible source of training data for models.
AI Cybersecurity Moves to the Front of the Agenda, Latent Space Says
Latent Space's daily AINews newsletter flags a trend: a run of fresh headlines has pushed AI cybersecurity to the top of the industry's 2026 agenda.
Kimi K3: The Largest Open AI Model Yet, Claimed to Match Opus 4.8 at Sonnet 5 Pricing
Moonshot AI has released Kimi K3, a 2.8-trillion-parameter MoE model (about 50B active) and the largest open model ever. It claims Opus 4.8-level quality at Sonnet 5 pricing.
Xaira unveils X-Cell, a 4.9B-parameter "virtual cell" for drug discovery
Xaira trained X-Cell, a 4.9B-parameter "virtual cell", on the largest genetic perturbation dataset yet, betting that causal models need causal data.
Poolside's Laguna S: a 118B MoE That Beats Thinky's Trillion-Parameter Open-Weights Model
AI lab Poolside has trained Laguna S, a 118B-parameter language model that, the company claims, outperforms the open Thinky model at roughly 1 trillion parameters.
Laguna S 2.1 vs DeepSeek V4: neolab promises cheaper than Flash and stronger than Pro
Lab neolab has unveiled Laguna S 2.1, claiming it is cheaper than DeepSeek V4 Flash and stronger than the flagship V4 Pro. The announcement includes no figures or benchmarks yet.
AI Engineer World's Fair: A Debate on Loops and a State-of-Engineering Report
The AIEWF conference concluded with a discussion on the role of loops in AI systems and presented a State of AI Engineering report, highlighting priorities for developers.
Lilian Weng: harness engineering and recursive self-improvement (35 articles)
Researcher Lilian Weng from Thinky published an analysis of 35 articles on harness engineering, emphasizing that explicit goal specification remains necessary even when integrating improvements into the model core.
Why AI infrastructure must evolve for agents — Akshat Bubna, CTO Modal
Modal's CTO explains how cloud infrastructure must change to support long-lived AI agents instead of short-lived functions.