Wednesday, July 1, 2026

Google DeepMind launches Nano Banana 2 Lite, its fastest and most cost-efficient Gemini image model, alongside Gemini Omni Flash for high-quality video and conversational editing. These models aim to scale creative workflows, reduce compute costs, and enhance real-time interaction.

shot-scraper 1.10 introduces the 'video' command, which accepts a YAML storyboard file defining a routine to run against a web application and uses Playwright to record a demo video. Written by GPT-5.5 in Codex Desktop, this feature enables coding agents to produce visual demos of their work. The article demonstrates creating demos for Datasette's bulk insert and CSV import features, detailing technical challenges and solutions.

Anthropic released Claude Sonnet 5, with performance close to Opus 4.8 at a lower price. However, a new tokenizer increases input tokens by ~30%, effectively raising costs. The model features a 1M token context window, 128K max output tokens, adaptive thinking enabled by default, and dropped temperature/top_p/top_k parameters. Its cyber capabilities are below Mythos 5, with safeguards matching Opus 4.7/4.8.

Anthropic has officially released Claude Sonnet 5, positioned as the most agentic Sonnet model to date, offering top-tier intelligence for coding and everyday professional work. The model demonstrates significant improvements in complex task execution, tool use, and autonomous decision-making, aiming to deliver a more efficient AI assistant experience for developers and enterprise users. Claude Sonnet 5 maintains the speed and cost advantages of the Sonnet series while achieving breakthroughs in reasoning accuracy and multi-step task planning.

Anthropic announces the redeployment of Claude Fable 5 starting July 1 after export controls are lifted. Updates include enhanced cybersecurity safeguards and a new industry jailbreak framework to improve model safety and controllability.

Anthropic launches Claude Science, a customizable AI workbench tailored for scientists. It integrates tools and packages commonly used by researchers, produces auditable artifacts, and offers flexible access to computing resources, aiming to accelerate scientific discovery and experimental analysis.

This paper proposes Reinforcement Learning with Metacognitive Feedback (RLMF), a paradigm that refines completion rankings based on the quality of model self-judgments, along with metacognitive data selection. Applied to faithful calibration, it aligns expressed with intrinsic uncertainty. Extensive experiments show RLMF achieves state-of-the-art calibration while preserving accuracy, surpassing standard RL by up to 63% and enhancing models' ability to assess and express capability limits.

LLM agents benefit from skill libraries for complex tasks, but selecting the right skill composition—subset, count, and order—remains a bottleneck. Existing methods either expose all skills or use retrieval, missing the structural nature. This paper formalizes structured skill composition and proposes SkillComposer, a constrained autoregressive decoder that jointly predicts the skill plan. Using a human-curated skill library for training, SkillComposer outperforms retrieval baselines in composition quality and downstream task success, matching the gold-skill upper bound with lower prompt-token cost.

IBM Research released ScarfBench on Hugging Face, a dedicated benchmark for evaluating AI agents on enterprise Java framework migration tasks. It covers diverse migration scenarios, measuring agents' ability to understand code, handle dependencies, and perform complex refactoring, providing a standardized evaluation for AI-assisted enterprise software modernization.

A paper on arXiv proves for the first time that Random Reshuffling (RR), a popular variant of shuffling SGD, dominates standard SGD in smooth convex optimization under any reasonable stepsize after any finite number of epochs. Previous theory required a stepsize threshold of O(1/n) and predicted worse rates for smaller epochs, contradicting empirical success. This work resolves a long-standing open question and provides rigorous theoretical justification for RR's superiority.

This paper proposes AutoBackSwap, which uses a secondary network to disentangle foreground and background, inpaint complete backgrounds, and augment training data by swapping various foregrounds with synthesized backgrounds. With only a few hundred patch-wise labeled samples, it can train the secondary network and automatically augment the full dataset. Even when no training sample breaks the spurious correlation, AutoBackSwap effectively reduces classifiers' reliance on backgrounds, consistently outperforming prior methods across tasks.

Semantic communication prioritizes task-relevant information for efficiency, but relay nodes introduce privacy risks: a relay can reliably infer semantics and reconstruct signals without accessing source data. This paper proposes an iterative adversarial training framework that accounts for a strong adaptive eavesdropper at the relay, optimizing representations to preserve high decoding performance at the intended receiver while degrading semantic inference at the relay. The approach significantly widens the semantic accuracy gap between legitimate receiver and eavesdropper across various channel conditions, achieving protection stealthily without compromising reconstruction fidelity.

This paper proposes converting user browser interaction trajectories into compact natural-language skills via skill distillation, enabling browser agents to read, retrieve, reuse, and compose these skills directly, overcoming the bottleneck of decision-making under incomplete information. The authors further organize distilled skills into a skill graph, allowing growth through consolidation rather than unbounded accumulation, suggesting that the scalability of browser agents may come more from the collective skills of Internet users than from manually designed tasks.

Conformal prediction provides a general framework for uncertainty quantification in predictive inference, but its application is often limited by computational cost. This paper further accelerates conformal prediction by incorporating approximate leave-one-out (ALO) estimators, establishing theoretical guarantees for asymptotic coverage and efficiency. Simulation results show that ALO-based methods achieve comparable coverage and efficiency to exact methods while significantly reducing runtime.

Acti launches a new keyboard for iOS and Android that embeds AI agents directly into the typing interface. Users can create custom AI shortcuts using natural language, enabling cross-app intelligent actions. The product aims to make the keyboard the next home for AI assistants, significantly boosting mobile task automation and interaction efficiency.

OpenClaw, a free and open-source agentic program, is now officially available on the Android and iOS app stores. It enables users to run localized AI agents on mobile devices for automation tasks, information retrieval, and content generation. This cross-platform launch marks the expansion of the open-source AI agent ecosystem from desktop to mobile, offering developers and general users a more accessible autonomous AI tool.

Venice AI announced a $65 million Series A funding round, reaching a unicorn valuation of over $1 billion. The privacy-first platform is already profitable with annualized run-rate revenues exceeding $70 million. CEO Erik Voorhees stated that the funds will be used to expand privacy-preserving AI services to meet growing demand for data security.

UK-based autonomous driving AI startup Wayve launches an $85 million employee tender offer at an $8.5 billion valuation. The move aims to retain key talent through liquidity incentives, reflecting a growing trend among AI startups using employee share buybacks as a strategic tool.

AI chip startup Etched announced it has booked $1 billion under contract for inference systems powered by its chip, sending its valuation to $5 billion. As a direct competitor to Nvidia, Etched has rapidly gained traction among cloud providers and enterprises with its chip architecture optimized for large-scale inference, signaling an accelerated shift in the AI chip landscape.

LuxEmo is a 21-hour conversational expressive speech corpus for Luxembourgish with 4 emotion categories, derived from RTL youth broadcasts via semi-automatic curation combining automated detection and human validation. The workflow includes VAD, denoising, language identification, ASR segmentation, emotion prediction, and lexical cues. Five expressive TTS systems (cross-lingual, multilingual, adaptation, non-parametric prosody transfer) are benchmarked using objective metrics and human evaluation.

EquiLibre Technologies, an AI lab founded by three former DeepMind researchers in Prague, is now valued at over $500 million. The team, which previously built a leading poker AI, is now applying game theory and reinforcement learning to generate profits for quant hedge funds.

The 'AI Compass' by bambamramfan is a political compass-style quiz that uses 29 questions about AI and AI ethics to classify test-takers into one of 30 archetypes. Simon Willison found it clever and thought-provoking.