Notes by Jin

Models & Research Meta's Brain2Qwerty v2 Narrows the Gap Between Non-Invasive Brain Decoding and Surgical Implants — Meta released Brain2Qwerty v2, an end-to-end pipeline that decodes typed sentences from non-invasive MEG recordings—no implant required. It reaches 61% word accuracy (up from ~8% for prior non-invasive methods) and 78%

Industry & Funding Mistral AI profile highlights rise as OpenAI rival — Mistral AI, known for offering some open-source models, has secured substantial investment since launching in 2023 as it pursues a mission of making cutting-edge AI broadly accessible. TechCrunch AI ↗ My takeaway: Mistral isn't winning on model quality

Models & Research Training a Single Transformer Layer with RL Can Match Full-Model Fine-Tuning — New research finds that reinforcement learning applied to just one transformer layer can achieve performance comparable to updating all model parameters, challenging the assumption that every layer must be trained during post-training. The work sheds light

Models & Research Anthropic details Fable 5 cybersecurity protections and a proposed jailbreak severity framework — Anthropic has restored access to Claude Fable 5 worldwide and shared additional information about the security measures guarding against cyber misuse, along with a draft framework for rating how severe a given jailbreak is. www.

Models & Research Redeploying Fable 5 — Anthropic is redeploying Claude Fable 5 on July 1 after the US lifted the export controls it imposed on June 12. Those controls followed a report that Fable 5's safeguards could be bypassed to find software vulnerabilities and, in one case, generate

Models & Research Introducing Claude Sonnet 5: Anthropic's Most Agentic Sonnet Yet — Anthropic released Claude Sonnet 5 on June 30, 2026, positioning it as its most agentic Sonnet model to date — capable of making plans, using tools like browsers and terminals, and running autonomously at a level that

Models & Research Self-Evolving World Models Aim to Improve Long-Horizon LLM Planning — Researchers propose WorldEvolver, a framework that lets LLM agents iteratively refine their internal world models to better predict the consequences of actions before executing them. The approach targets a core weakness in agentic AI where unreliable foresight can

Models & Research Palantir Deploys NVIDIA Nemotron Open Models for U.S. Government Agencies — Palantir has launched a new AI engine built on NVIDIA's Nemotron open-source models, designed to operate within secure environments for U.S. federal agencies. The collaboration highlights how open-source AI can serve sensitive government

Models & Research OpenAI Previews the GPT-5.6 Model Family (Sol, Terra, Luna) Behind a Phased, Government-Coordinated Release — OpenAI has begun a limited preview of its GPT-5.6 series: flagship Sol, balanced everyday model Terra, and fast, low-cost Luna. Sol is framed as its strongest model yet, paired with its

Industry & Funding Tech giants race to build custom AI chips, challenging Nvidia's dominance — OpenAI has revealed plans for a custom inference chip called Jalapeño, developed in partnership with Broadcom, joining companies like Google, Apple, and SpaceX in a broader industry trend of building proprietary silicon to reduce

Policy & Society White House Asks OpenAI to Delay New Model Release Over Safety Concerns — The Trump administration has reportedly pressured OpenAI to limit the rollout of its latest model, GPT-5.6, to a small group of partners rather than making it broadly available. The request reflects growing government involvement

Models & Research Real-Time Voice AI Systems Process Words But Miss Vocal Delivery Cues — A study testing four leading production voice AI platforms found that while they accurately handle spoken words, they largely fail to interpret meaning conveyed through tone and other vocal characteristics. The gap reveals a significant limitation

Models & Research GPT-5 Assists Immunologist in Cracking a Long-Standing Scientific Mystery — An immunologist used GPT-5 Pro to resolve a three-year-old puzzle about T cell behavior, a finding that may open new avenues for cancer and autoimmune disease research. The case highlights how frontier AI models can accelerate biological discovery.

Industry & Funding Google DeepMind Commits $75M to AI Filmmaking With A24 Partnership — Google DeepMind is partnering with acclaimed studio A24 in a $75 million deal aimed at developing artificial intelligence tools designed specifically for the film industry. TechCrunch AI ↗ My takeaway: A24's know-how could be the critical

Models & Research GLM-5.2: An Open 1M-Context Model Built for Long-Horizon Coding — GLM-5.2 is Z.ai's new flagship model, built around a "solid" 1M-token context that stays usable across long coding-agent runs rather than just accepting more tokens. On Z.ai's own

Tools & Open Source "In the Weights" Lets You Check Your AI Training Data Footprint — Former OpenAI employees Thomas Dimson and Joey Flynn built a tool called In the Weights that queries multiple AI models to measure how well they can recall a given person, assigning a strength

Industry & Funding Reliance's Ambani Pushes AI Integration Across Half a Billion Telecom Users — Indian conglomerate Reliance is embedding AI capabilities throughout its telecommunications infrastructure, which serves more than 500 million customers, with ambitions to make the technology present in calls, applications, and households. The initiative represents one

Models & Research Fine-Tuned LLMs May Pattern-Match Rather Than Reason About Vulnerabilities — A new benchmarking framework built from hundreds of manually curated Linux kernel samples suggests that LLMs performing well on vulnerability detection may be exploiting data patterns rather than genuinely reasoning about security flaws. The findings raise questions about

Industry & Funding HSBC Deepens Multi-Year AI Collaboration with Google Cloud — The banking group announced a broad, multi-year agreement with Google Cloud to roll out artificial intelligence capabilities spanning wealth management, financial crime detection, and internal decision-making across its global footprint. The deal was revealed at Google Cloud Summit London.

Industry & Funding Plaud Crosses $100M ARR Milestone With AI Notetaker Hardware and Software — Hardware-software startup Plaud has reached over $100 million in annual recurring software revenue after shipping more than two million of its AI-powered notetaking devices. The milestone comes as competition in the AI meeting assistant space intensifies.

Policy & Society EU Releases AI Content Labelling Guidance Ahead of August Deadline — The European Commission has issued a voluntary Code of Practice outlining concrete steps companies should take to meet transparency obligations that become legally binding across the bloc this August. The playbook is designed to help organizations prepare