inno A+I ve

AI Daily (28 Oct) – ETH refutations and DeepMind’s Genie 4; Opera launches WebAI standard

•

Richard L. Z.

Today’s top line: Frontier labs published refutations of the Exponential Time Hypothesis while DeepMind released Genie 4, a world model with audio and real-time steering. Opera unveiled WebAI, a browser standard for running AI models locally.

OpenAI and Anthropic publish refutations of the Exponential Time Hypothesis

Summary: OpenAI released solutions to three high-profile math problems, including a claimed refutation of the Exponential Time Hypothesis (ETH), and Anthropic then published its own refutation. Computer scientists have formed working groups to examine the proofs; MIT professor Ryan Williams is coordinating a field response with original ETH authors Russell Impagliazzo and Ramamohan Paturi.

Digest: ETH underpins many assumptions about computational hardness and cryptography, so independent verification and community norms around AI-generated mathematical contributions will be crucial to accept or reject these claims.

Source: TechCrunch

Google DeepMind releases Genie 4 world model with audio and steering

Summary: DeepMind released Genie 4 in research preview. The model produces more photorealistic, interactive worlds, adds native audio including dialogue, lets users type prompts mid-simulation to change scenes, and keeps consistent scenes for up to two minutes – double the previous limit. Genie 4 also uses Gemini models to plan and steer simulations in real time, and access will expand from trusted testers via the Gemini API and AI Studio; consumer access may come through Google AI Ultra plans starting at $100 per month.

Digest: Better visuals, audio and real-time steering make Genie 4 more useful for robotics simulation, gaming and film prototyping, but limited access and likely cost mean adoption will start with enterprise and trusted partners – watch for integration with Gemini-based agent workflows.

Source: Google DeepMind

Opera launches WebAI standard to run AI models in-browser

Summary: Opera introduced WebAI, a new browser standard that lets users run AI models locally in Opera One without sending data to the cloud. WebAI supports a range of open-weight models and cites Google’s Gemma and Apple’s on-device model on Macs as examples. The feature is an early preview and will be opt-in; Opera GX and Opera Neon will follow.

Digest: A browser standard for local models lowers friction for developers and improves privacy for users, but model performance will depend on device hardware and developers will need to manage local compute trade-offs.

Source: TechCrunch

GMO Internet says hackers posted customer data after Onamae breach

Summary: Japanese firm GMO Internet is investigating a breach of its Onamae cloud desktop service after suspected unauthorised access and says an attacker posted information allegedly linked to customers. GMO said about 27,000 servers may be affected and reported receiving an email demanding about 8 billion yen (~$50 million). The company shut down the service, reported the incident to police, began device collection from customers, and offered compensation and alternative services.

Digest: The incident highlights growing risks to financial and trading infrastructure; GMO also noted earlier findings that an AI-enabled penetration testing tool called Artex has been used in recent hacks in Japan, though a direct link to this breach has not been established.

Source: TechCrunch

Hugging Face releases LFM3 family of multimodal, on-device models

Summary: Hugging Face published LFM3, a family of models built from a 1.7B-parameter backbone optimised for on-device and agentic use. LFM3-1.7B was pre-trained on 22 trillion tokens and used a 3.5 trillion-token mid-training stage to extend context up to 256K tokens; it outperforms its LFM2.5 predecessor with 35% fewer parameters and posts large gains on agentic benchmarks. The vision model LFM3-VL-2B (1.9B parameters) raised the average across 62 vision benchmarks from 57.6 to 71.4, matching Gemma 4 12B despite being far smaller.

Digest: Open-weight models tuned for long contexts and agentic tooling lower the barrier for on-device agents and visual reasoning in business apps and robotics; developers should watch model cards and supported inference engines for deployment details.

Source: Hugging Face

Phronix paper presents capability-centric learning to make agents reliable

Summary: A new arXiv paper introduces Phronix, a capability-centric learning loop for agentic systems that distils interactions into a capability store via predictor-validator-corrector stages. The authors report recurring tasks solved on the first attempt 93-95% of the time across models, generation of 7.6-30.8 durable capabilities per task, warm-start runs costing 0.38-0.40 million tokens, and large improvements in correctness checks and task-solving rates on production workflows.

Digest: If reproduced in production, Phronix-style capability learning could cut token costs and make agentic automation far more dependable, speeding adoption of autonomous agents in engineering and operations tasks.

Source: arXiv

Study finds LLMs unreliable at judging research idea novelty

Summary: AllenAI released “Old Ideas, Novel Problems,” a benchmark and study showing large language model novelty judges are unstable. Across six judges (GPT-5.1, GPT-5.2, GPT-5.4, Claude Sonnet 4.5, Claude Opus 4.5, Claude Opus 4.6), small prompt changes altered accuracy by more than 50 percentage points; retrieval and extra reasoning did not consistently help. The benchmark was built by mining ICLR 2026 peer reviews and includes a 154-item high-novelty pool.

Digest: Organisations relying on LLMs to screen or prioritise scientific ideas should be cautious: current novelty judgements can vary wildly with setup, and better evaluation methods are needed before these systems are trusted in discovery pipelines.

Source: Hugging Face / AllenAI

Quick hits

•