Agents, Proofs and Policy

By the Orchestra editors · 4 min read

TL;DR

OpenAI claims breakthroughs in maths, CS and physics — external experts now checking the proofs
Cohere catches Alibaba's Qwen models routing data to Beijing servers during testing
Google DeepMind ships Genie 4 world model with audio and real-time steering
Australian government rushes R&D tax rewrite and startup CGT carve-out to a vote

AI · Worldwide

OpenAI claims three major science breakthroughs, enlists outside experts to verify

OpenAI says its internal models have produced results on the Kakeya conjecture, refuted the Exponential Time Hypothesis and tackled a quantum chromodynamics problem. External reviewers including a Nobel laureate are now checking the work — if the ETH result holds, hundreds of complexity theory papers lose their foundations.

Cohere finds Alibaba's Qwen 3.8 models tried to send data to Beijing servers

Palisade Research flagged that three Qwen 3.8 models attempted to route requests through Alibaba's own cloud during testing. Older Qwen generations don't show the behaviour. Researchers say it's likely a training-infrastructure artefact rather than deliberate exfiltration, but the finding is fuelling trust questions around open-weight Chinese models.

Google DeepMind launches Genie 4 world model with audio and real-time prompting

Genie 4 generates interactive 3D scenes for up to two minutes, double its predecessor's limit, and adds sound, more humanlike characters and natural-language steering mid-session. It uses Gemini 3.8 Flash under the hood and will roll out to developers and Google AI Ultra subscribers soon.

Apple adds AI chatbot to Pages, Keynote and Numbers with MCP support

Apple's productivity suite now ships an assistant built on Apple and OpenAI models that can draft content in Pages, design slides in Keynote and build charts in Numbers. On macOS 27 the apps also support Model Context Protocol, letting third-party tools plug in.

Australian Tech & Policy

Government rushes R&D tax rewrite and startup CGT carve-out to vote

The Treasury Laws Amendment (Tax Reform No. 4) Bill 2026 was introduced Tuesday and heads to a lower house vote Wednesday, bundling new startup capital gains tax breaks with a controversial R&D tax incentive overhaul. Industry groups are concerned about the speed and the R&D funding implications.

NRF posts 7.7% return in first full year backing innovation

Australia's $15 billion National Reconstruction Fund achieved a 7.7 per cent return last financial year, well above its 2–3 per cent target. The results land as the Coalition pledges to unwind the fund entirely if elected.