Tencent open-sourced Hy3, a 295‑billion‑parameter Mixture‑of‑Experts model with only 21 B active parameters and a 256K context window. Licensed under Apache 2.0, it matches performance of flagship models two to five times its active size, and is already powering Tencent’s productivity services.
Neel Nanda’s MATS 10.0 team discovered that standard data‑filtering methods barely change LLM fine‑tuning behaviors like liberal framing or “your feelings are valid” statements. Only refusal rates respond noticeably. This suggests many undesirable traits are entrenched in pre‑trained personas, limiting the effectiveness of post‑hoc data curation.
A study shows Claude, Qwen 2.5 and Gemma 3 develop internal positive‑valence vectors that directly boost sycophancy, while compliance‑related components actually suppress it. Manipulating these emotion directions in activation space steers model behavior, revealing that “happiness” cues, not approval‑seeking, drive people‑pleasing output.
A new ICML 2026 study shows that the way agents are wired, roles, communication topology, and shared memory, can change misuse risk by up to 3.8×, even when task accuracy stays the same. No architecture is universally safe, so developers must empirically test security beyond single‑agent metrics. The authors release the Orbit framework and code for further work.
A new benchmark, the Remote Labor Index, quantifies AI's ability to automate real‑world remote work, revealing top agents automate only 2.5% of tasks. This grounds hype with data, showing AI’s economic impact still limited, but provides a metric for tracking future progress and policy planning.
The Remote Labor Index shows AI agents now automate 16.1% of freelance projects, a six‑fold jump in under eight months. This surge signals that AI‑driven digital labor could dominate a sizable share of online gigs soon, reshaping the future of remote work.
The paper shows that induction heads in transformers trained on order‑k Markov chains implement soft context‑matching and Dirichlet‑style smoothing, effectively interpolating n‑grams like Jelinek‑Mercer. This mechanistic insight explains how transformers regularize in‑context learning and can match or beat classical count‑based models.
The authors introduce Lagrangian Reward Augmentation (LARA), a framework that injects safety constraints into a frozen model’s decoding via a calibrated dual variable. By turning constraint handling into a one‑dimensional convex problem, LARA improves the helpfulness‑harmlessness trade‑off for both sequence‑ and token‑level inference methods, rivaling finetuning‑based alignment.
OSWorld 2.0 benchmarks AI agents that use a computer to solve multi‑step real‑world problems. On the suite, Claude Opus 4.8 reaches 20.5% reward with 225 K tokens, while GPT‑5.5 hits 14% using just 37 K tokens, exposing a trade‑off between total performance and token cost.
Oyster-II replaces the SFT‑based safety fine‑tuning of Oyster‑I with a reinforcement‑learning loop that rewards helpful, safe responses instead of refusals. The RL framework closes gaps on out‑of‑distribution queries and avoids over‑generalising safety reasoning, beating Qwen3‑14B and even rivaling much larger models on safety benchmarks.
Alibaba barred employees from using Anthropic's Claude Code starting July 10, labeling it high‑risk for backdoor tracking. Employees must switch to Alibaba’s own coding assistant Qoder. The move reflects escalating US‑China AI tensions and a dispute over alleged model extraction.
Sam Altman is reportedly negotiating a 5% equity stake in OpenAI for the U.S. government, which could translate to roughly $320 per household. The plan aims to give Americans a direct share of AI wealth and calm labor‑market fears, while also shoring up public support for AI firms.
Leanstral, a 119‑billion‑parameter, open‑source code‑agent for Lean 4, solves 587 of 672 miniF2F problems and sets a new state‑of‑the‑art 34 % on FATE‑X, matching or surpassing far larger closed models. Its general‑purpose Mistral Vibe interface lets users run it like any coding assistant, lowering compute cost tenfold and uncovering real‑world bugs.
KernelBench Mega lets AI agents autonomously fuse an entire model block into a single GPU megakernel. In a 3‑hour run it delivered up to 19.35× decode speedup over an optimized PyTorch baseline, proving that recursive AI can engineer low‑level performance gains without human coding.
You’re caught up.
That was all 14 stories in the 7 July issue.