Topic
Everything filed under LLMs, newest first.
RSS · JSON · All topics
Introducing Ripple-65B Preview, a Frontier Open-Weight Model
Ricursive Intelligence introduces Ripple-65B Preview, a language model trained from scratch to power its chip design platform, which it says sets a new U.S. open-weight state of the art on agentic coding, with notes on architecture, training efficiency and compute.
4 min · 1,003 words
Introducing Kinference: An Open Source Inference Control Plane for vLLM and Kubernetes
vLLM and collaborators introduce Kinference, an open-source inference control plane for Kubernetes that adds token rate limiting, multi-tenant controls, and operational primitives on top of vLLM serving.
4 min · 1,034 words
LLM Judges for Identity Resolution at 4x Lower Cost
We wanted to test whether LLMs and decision models, specifically Jev and Luna, could solve a hard problem that we encounter regularly (grading identity resolution results). We found that while both models performed okay out of the box (88.8–97.3% accuracy), implementing a hybrid model that combines Jev + Opus achieves frontier accuracy at > 4x lower costs. - Luna performs better than Jev out of the box, but the gap disappears on harder cases. - Jev's confidence is better calibrated. Most of its mistakes come with low confidence its right. Most of Luna's mistakes come with…
11 min · 2,468 words
Investigating vault particles with Claude
Anthropic’s biology lab shares the first 3D structure of a vault-like particle found in prokaryotic cells, uncovered with Claude-assisted structural biology workflows.
11 min · 2,538 words
Tencent Launches Hy4, Joining the Ranks of the World's Leading Open-Source AI Models
Tencent announces Hy4, its latest-generation large language model among leading open-source AI models, with plans to release weights and system cards.
5 min · 1,139 words
Cognition introduces SWE-2 Flash, a U.S.-trained coding model that adds a mid-training stage before RL and aims for high intelligence at low cost and high speed.
3 min · 678 words
TIS-100 Bench: LLMs Push the Pareto Frontier in Programming Puzzle Games
Wincon gave 16 LLMs the puzzle game TIS-100 and its manual, withheld the community leaderboard, and let them keep optimizing. The models found 133 new Pareto frontier points on 18 of 55 puzzles, including 3 outright cycle records, and players quickly improved on them.
17 min · 3,982 words
How We Optimized Browser Use for Deep Agents
LangChain explains how it made browser use efficient for Managed Deep Agents: context-tuned Stagehand tools, selective observation, and a human handoff path so agents can browse without drowning in page noise.
8 min · 1,820 words
Don't Call It a Decision Model
TypeSafe AI explains why it calls Jev a System One model rather than a "decision model": Jev makes atomic judgments while surrounding code makes decisions, and the post shows how to design systems around that split.
9 min · 2,029 words
Qwen 3.8 27B on my laptop can interpret model weights, too
Simon Willison tries an example from the 'Sparks of Weight-Space Programming' paper and finds Qwen 3.8 27B running locally on his laptop can also interpret raw neural network weights written out as text, while arguing the wider commentary that source code is over goes too far.
5 min · 1,241 words
New context engineering challenges in late 2026
Harrison Chase surveys late-2026 context engineering challenges for agent builders: longer horizons, tool noise, memory tradeoffs, and why harness design now matters as much as the underlying model.
4 min · 892 words
Nace.AI releases Drex for Finance, a decision model for fintech fraud, credit and filings that returns option probabilities instead of free text, scoring 77.3 on its finance index versus 63.6 for Jev 1.14.
5 min · 1,241 words
Klang Research sent the same 3,000 sentences through twelve frontier AI models and found every one charges 30% or more extra tokens for Swedish, Norwegian, Danish and Finnish than for English, explaining where the extra tokens come from.
5 min · 1,037 words
The shelf life of LLM findings
Stanford's Owen Queen and James Zou re-ran 50 widely cited findings about language models, from sycophancy and jailbreaks to fabricated citations and long-context limits, across 20 models released July 2024 to September 2026, finding that within 15 months most claims had expired or substantially weakened.
28 min · 6,536 words
so i guess everyone uses llms to think now
Security researcher eva explains why she never uses AI for the research on her blog: tinkering is the fun, she can't take credit for an LLM's work, and she argues AI firms use hyped vulnerability disclosures, PoC videos and hash drops (php2shell, 'tailworm') as marketing.
8 min · 1,760 words
Appraisal by AI: Model-driven victim prioritization in the VOLTA infostealer
Cisco Talos analyses VOLTA, a self-deleting infostealer that, after stealing credentials, sessions and crypto wallets, sends a structured evidence bundle to an LLM to rank the victim's assets by monetization value, failing over across a three-model chain when a model refuses.
7 min · 1,556 words
A year after his last stock-take, Salesforce AI Research engineer Ben Stolovitz describes how his AI use has changed: his personal AI policy, heavy Claude Code use for coding, research and search, summarization, transcription and voice, writing, and art, plus the subscriptions he keeps.
21 min · 4,863 words
Post training a model to explore Demoscenes
datavorous describes post-training Demoscener-7B on a dataset of about 25k tiny demoscene programs (dweets) to explore how sizecoded visual effects arise from unfinished code, with examples and released model and dataset on Hugging Face.
6 min · 1,485 words
Ryan Madden reports that GPT-5.6 Sol, running in Codex at a low thinking level, completed a peaceful (no biters) Factorio run to rocket launch in 13.5 hours of game time via a programmatic game library, and describes where the model excelled and where late-game scaling broke down.
2 min · 364 words
Mistral Large 4 (Le Chonk) is now open weight
Mistral AI releases the weights of an improved Mistral Large 4 (Le Chonk), its 1-trillion-parameter natively multimodal model, after further reinforcement learning since the public preview. The post covers gains in agentic coding and defensive cybersecurity benchmarks, three reasoning modes, enterprise and sovereignty use cases, partner quotes, the Mistral Ecosystem Open-Weight license, and new API pricing.
15 min · 3,342 words