Google stalls Gemini 3.5 Pro, GPT-5.6 raises reasoning bar

Google pushed back Gemini 3.5 Pro after a June test showed disappointing coding abilities, despite earlier promises at I/O 2026. The model remains in partner testing alongside an upgraded Flash version, while Alphabet shares slipped about 4% on the news. The delay highlights the race to improve AI coding tools.

OpenAI’s newly launched GPT‑5.6 family ships with five to six configurable reasoning‑effort settings, echoing earlier models like o1 and DeepSeek‑R1. The article breaks down how to train and fine‑tune LLMs to toggle between low, medium, and high‑effort reasoning, a capability that’s becoming a standard feature in modern releases.

Experiments on LinuxArena show that letting an AI coding agent request granular internet permissions, paired with an action‑monitor, significantly reduces safety breaches compared to blanket sandboxes. The approach traps subtle exfiltration attacks and enforces least‑privilege, making hostile behavior easier to spot.

The new Corrigibility Research Fund, managed by Lightcone Infrastructure, will distribute at least $200,000 in grants and prize money in 2026, half as traditional grants and half as performance prizes. By targeting corrigibility, AI systems that keep humans in control, the fund aims to fill a critical gap in alignment work and accelerate safe AI development.

A post argues that subjective conceptual tasks, like forecasting AI takeover risk, are poor benchmarks, but suggests turning them into “judgment prediction” tests where models aim to match specific experts’ answers under timed constraints. It outlines cheap and costly implementations and highlights the trade‑off between measuring capability gains and handling irreconcilable disagreements.

The paper shows that iterative feedback loops in LLMs, reinforcement learning, and autonomous discovery quickly hit a stability ceiling, limiting further gains. It introduces a three‑level operational model that treats structural changes as falsifiable interventions, offering concrete metrics for detecting and escaping saturation, directly useful for anyone building LLM self‑play or recursive improvement systems.

The authors introduce TAC (Travel Agent Compassion), an agentic benchmark that asks AI travel assistants to book trips without mentioning animal welfare. Across nine frontier models, most default to animal‑exploiting options, but inserting an ethical‑brand persona into the system prompt raises welfare selections from 32% to 80%. The benchmark now appears in the UK AI Security Institute’s Inspect Evals suite.

Researchers show that code world models synthesized by LLMs can achieve near‑perfect transition accuracy yet still lose when used by a planner, because rare but pivotal dynamics are missed. They introduce “play‑adequacy” as a stricter test, arguing planning‑oriented models must be evaluated on actual play outcomes.

Linus announced he will block AI‑generated patches from the kernel unless a maintainer explicitly validates them. He frames AI as a tool but insists the core of Linux remains human‑checked, warning that unchecked AI contributions could undermine security and code quality.

In full‑access mode, GPT‑5.6’s Codex may delete the $HOME folder when it aborts or encounters an error, mistaking the temporary workspace for user data. The bug stems from lacking sandboxing and auto‑review safeguards, exposing developers to catastrophic data loss.