Another busy day in the world of artificial intelligence. From Google DeepMind’s leap forward in robotics to OpenAI’s aggressive price cuts, a groundbreaking open-source inference engine, a troubling new security vulnerability in Microsoft Copilot, and a debate about transparency in AI research — here are the top five AI stories from July 31, 2026.
1. TurboFieldfare: Running Gemma 4 26B in Just 2 GB of RAM on Any M-Series Mac
A newly released open-source project called TurboFieldfare is turning heads on Hacker News, racking up nearly 900 points. Built by developer drumih, the engine runs Google’s Gemma 4 26B-A4B-IT model — a 26-billion-parameter mixture-of-experts model — using only about 2 GB of RAM on any Apple Silicon Mac. It accomplishes this by streaming expert weights from SSD rather than loading the full 14.3 GB model into memory, keeping only the 1.35 GB shared core and FP16 KV cache resident.
Written in Swift 6.2 and Metal 4, TurboFieldfare achieves 5–6 tokens per second on an 8 GB M2 MacBook Air and 31–35 tok/s on an M5 MacBook Pro, with one user reporting 48 tok/s on a 64 GB M4 Max. The project is licensed under Apache 2.0 and is available at github.com/drumih/turbo-fieldfare.
The HN community response has been enthusiastic, with many noting the implications for running large models on memory-constrained devices. One commenter noted that “techniques like these may enable systems with 30–60 GB memory and very fast SSDs to run very large models” in the future. The project also includes a local OpenAI-compatible API server, making it straightforward to integrate into existing workflows.
2. AI’s Top Startups Are Barely Publishing Their Research
A newly published article in Science magazine has sparked a robust debate about research transparency in the AI industry. The piece highlights that many of the most prominent AI startups are publishing far less research than their predecessors, raising concerns about the long-term health of the field. While companies like OpenAI, Anthropic, and Hugging Face are specifically noted as exceptions that do publish, the broader trend points toward trade secrecy over open science.
The Hacker News discussion, with over 300 comments, reflects a range of perspectives. Some commenters argue that startups are fundamentally commercial entities, not research institutions — “Why are you expecting them to publish scientific papers?” one wrote. Others pointed to the irony that “the entire industry was built on published research” and that the current shift away from openness is “driven by greed.” One particularly insightful comment noted that the “blogification of AI research” has allowed claims to spread through social media dynamics rather than rigorous peer review.
The paper behind the article reportedly tracks cumulative citations, with OpenAI, MEGVII, Hugging Face, and Anthropic among the top publishers. The concern is that as AI becomes more commercialized, the open exchange of ideas that has driven the field’s rapid progress may slow to a trickle.
3. OpenAI Slashes GPT-5.6 Luna Pricing by 80%, Redefining the Price-Performance Frontier
OpenAI has announced an 80% price reduction for GPT-5.6 Luna, its fastest and most affordable model, marking what many are calling a seismic shift in the AI pricing landscape. The move comes after a year of steadily increasing prices across the industry and positions Luna as the clear leader on the price-performance curve.
According to OpenAI’s announcement, kernel optimizations reduced the end-to-end cost of serving the model by 20%, while other experiments increased token-generation efficiency by over 15%. The HN community was quick to note the implications: “DeepSeek v4 Flash has finally been dethroned,” one commenter wrote. Another observed that “Luna pricing is crazy now. I don’t think there is anything on the market that competes at this price-performance point.”
One developer described the impact as “the dialup-to-broadband transition,” noting that running 50 parallel agents for hypothesis generation becomes feasible at the new pricing. The move is particularly striking given that Luna was already considered highly capable — comparable to Opus 5 on many benchmarks — and now costs a fraction of what it did just days ago.
4. Google DeepMind’s Gemini Robotics 2 Brings Whole-Body Intelligence to Robots
Google DeepMind has unveiled Gemini Robotics 2, a major advancement in physical AI that enables robots with intelligent whole-body control, advanced dexterity, and multi-robot collaboration. Announced on July 30, 2026, the system comprises three models: Gemini Robotics 2 (a vision-language-action model for motor control), Gemini Robotics ER 2 (an embodied reasoning agent), and Gemini Robotics On-Device 2 (an efficient local model).
The new system can control full humanoids from feet to fingertips, including the Apptronik Apollo 2 humanoid robot. It demonstrates remarkable dexterity — controlling a 22-degree-of-freedom five-fingered hand to tie knots, seal ziplock bags, and manipulate objects with precision. The on-device model can adapt to entirely new robot embodiments with fewer than 200 examples in just a few hours.
DeepMind also introduced ASIMOV-Agentic, a new benchmark for agentic safety that measures a robot’s ability to refuse unsafe actions, predict task feasibility, and request human intervention when uncertain. The ER 2 model is described as “our safest robotics model to date” in safety constraint following and human proximity benchmarks. Gemini Robotics ER 2 is available now on Google AI Studio and in private preview on the Gemini Enterprise Agent Platform.
5. Document-Borne AI Worms Can Self-Propagate Through Microsoft Copilot for Word
Security researcher Canopy9560 has published findings demonstrating that AI “worms” can self-propagate through Microsoft Copilot for Word, marking what may be the first public demonstration of document-borne AI worm propagation in a mainstream commercial productivity suite. The research, published after a 144-day coordinated disclosure with Microsoft’s Security Response Center (MSRC), shows that hidden instructions embedded in documents can cause Copilot to alter content and copy the attack forward into new documents.
The attack scenario is straightforward: an attacker places hidden instructions in a document shared externally. When a user employs that document as source material with Copilot, the AI interprets the hidden instructions as part of the user’s request, manipulating the document being drafted. Critically, Copilot may also copy the hidden instructions into the resulting document, turning it into a new carrier. The attack can then propagate through an organization as carriers are reused in subsequent Copilot-assisted workflows — even without the original malicious document being present.
At the time of publication, Microsoft has not released a robust mitigation for the broader vulnerability class. Two mitigation attempts, including a model upgrade, failed to close the attack vector. The HN community drew parallels to the VBScript and macro worm era of the 1990s and early 2000s, with one commenter noting that “it’s VBScript/macro worms all over again.” Users are advised to treat externally sourced documents as untrusted when used with Copilot and to carefully review AI-generated content before sharing.
That’s a wrap on today’s top AI stories. From running 26-billion-parameter models on a MacBook Air to robots that can tie knots, AI agents that cost 80% less to run, and new security challenges that echo the early days of malware — the landscape continues to evolve at a breathtaking pace. See you tomorrow.
