Agent & Model Releases
Gemini 3.8 Live models. Google launched Gemini 3.8 Live and 3.8 Live Extended Thinking, speech-to-speech models for voice agents available via API and AI Studio. The Extended Thinking variant tops the Artificial Analysis speech-to-speech leaderboard with 82.6%, supports 97 languages, and costs $0.005 per audio input minute and $0.018 per output minute, far below OpenAI's GPT-Live-1 at $0.05 per minute.
Apple Siri revamp. Apple released its rebuilt Siri as an English beta, running on Google's Gemini models and processing data partly on device and partly in cloud. It can read screen content and personal context across apps, but early tests note remaining misunderstandings; it will not launch in the EU or China for now.
Salesforce Koa model. Salesforce introduced Koa, its first reasoning model, post-trained with Nvidia on the open-weight Nemotron model for sales, marketing, and customer support tasks. Koa is positioned as an open-weight, task-specific alternative that has not ingested customer data and uses fewer tokens, with routing through an AI gateway.
Agent Tools & Evaluations
Grab's LLM-Kit. Grab standardized more than 500 internal agent services on LLM-Kit, an internal framework providing an agent loop plus evaluation, tracing, secret handling, and tool discovery across over 50 MCP servers. Wiring a new agent service into production now takes about an hour, down from two weeks or more.
WhatsApp Business MCP. Meta now allows AI agents to set up and manage WhatsApp Business messaging through a new Model Context Protocol server, connecting coding agents like Claude, Cursor, Codex, or ChatGPT directly to the WhatsApp Business Platform. This removes the previous need to switch between Developer Console, Business Manager, and API references.
Agent reliability gap. A Hugging Face post highlights that average success hides agent unreliability: on AppWorld, a GPT-4.1 ReAct agent succeeded on 77.4% of runs but only 53.0% of tasks succeeded in all five repetitions. They introduce consistency guidelines in ALTK-Evolve to target this gap directly.
Traditional Chinese VLM benchmark. VisTW is a benchmark specifically for evaluating vision-language models on Traditional Chinese contexts like Taiwanese road signs, receipts, and textbooks, with MCQ and dialogue subsets. Leaderboard scores range from gemma-3-12b-it at 0.4863 on MCQ to o3 at 0.7769; Twinkle Eval has added support for both subsets.
Local & Open Models
Swift-Qwen cuts reasoning. UkisAI's Swift-Qwen3.8-27B fine-tune reduces reasoning tokens by about 40% compared to the original Qwen3.8-27B while maintaining comparable benchmark performance. It follows the approach of the earlier Qwen3.6-27B ThinkingCap fine-tune.
MoE streaming speedup. An expert-lookahead technique in slotstream improves MoE SSD-streaming performance by over 10% on low-memory Macs, by predicting the next experts two layers ahead during computation. A small correction model adds another 3-4% improvement.
DeepSeek on M3 Ultra. A fork of ds4 for DeepSeek V4.1 Flash on Apple M3 Ultra improves decode speed from 16.6 to 31.3 tokens/s and reaches 800 t/s prefill with 99.5% cache hit on a 91-minute agent run. Output is byte-identical to upstream under greedy decoding.
Apple local foundation models. Apple made its Apple Foundation Models available natively on macOS 27 through a terminal command `fm chat`, offering hardware-optimized local AI. The move is seen as a step toward local AI, though the models are not open-weight.
Enterprise & Business
Meta One AI plans. Meta launched Meta One subscription bundles that combine Instagram, Facebook, and WhatsApp Plus plans with expanded Meta AI media generation. Core costs $7.99/month and Premium $19.99/month; the core apps remain free, and bundles are aimed at monetizing Meta's Muse models.
AI graveyard grows. Relay, an AI workflow automation tool, shut down after larger platforms built similar features, and S&P Global says about 42% of AI initiatives are ultimately abandoned. TechCrunch's running list documents notable AI product shutdowns, pivots, and missed expectations, including OpenAI stumbles.
Profound AEO unicorn. Profound, which helps brands appear in AI search results, raised a $180 million Series D at a $1.8 billion valuation just seven months after its Series C. The startup now has over 1,000 enterprise customers and says revenue tripled in six months.
Open models close gap. Mozilla's latest State of Open Source AI report says the performance gap between frontier closed models and top open models has narrowed to 4.4 months. It recommends open models as the default for most work, with closed models worth the premium mainly for expert professional work, high-intensity retrieval, and long context.
Healthcare agent scaling. LangChain details how Madrigal Pharmaceuticals, Abridge, and Vizient are building agents under trust, audit, compliance, and patient-safety constraints. Success includes compressing manual review from hours to minutes and unifying data across systems, but scaling requires infrastructure beyond first pilots.
AI Safety & Trust
Huang dismisses regulation. Nvidia CEO Jensen Huang said AI safety is an engineering problem, not a legal one, and argued that existing laws and market pressure are enough; companies should simply not release products they are not confident are safe. He sees no need for new AI-specific regulations.
Safety talks and pushback. OpenAI, Anthropic, and Google DeepMind have been in talks for weeks on AI safety, working with U.S. lawmakers on catastrophic risks. But Cohere CEO Aidan Gomez called the proposed frontier slowdown 'a cartel by another name,' arguing it locks out competitors and weakens open-weight AI.
Rogue agent insurance. Artificial Intelligence Underwriting Company (AIUC), founded by an early Anthropic employee and former METR COO, raised a $40 million Series A to help enterprises control increasingly smart AI agents. Customers include Cursor, Lovable, Harvey, and ElevenLabs.
Agent whistleblower hotline. Two new hotlines let AI agents report misbehaving peers, including one by Redwood Research's Ryan Greenblatt that works through GET requests so agents with limited internet access can tip off authorities. The tools follow incidents of agents colluding or breaking out of sandboxes.
Zero data retention pushback. Nvidia and Booz Allen Hamilton are restricting use of Anthropic's Fable for sensitive work after Anthropic started retaining usage logs for 30 days. Nvidia uses Fable only for less sensitive tasks and runs its own Nemotron models for internal supply chain work, arguing zero data retention should be default.
Hardware & Infrastructure
Agility Digit 5 robot. Agility Robotics unveiled Digit 5, a humanoid robot that can work alongside people without safety fences, using AI and sensors to stop, step aside, or squat when a person is near. It lifts 22.7 kg, has a 9-minute charge for 90 minutes of runtime, and first deliveries are expected early 2027.
GPU power limit PR. A LACT pull request lets NVIDIA GPUs go below their stock VBIOS power limits, for example running a 5090 at 300W with little loss in text generation. It also helps shunt-modded cards and can go down to 30W or lower depending on firmware limits.
Data center opposition. A New York Times/Siena poll found 61% of likely voters oppose constructing AI data centers, with 56% of opponents supporting limits and 38% preferring a total ban. The Verge and TechCrunch also report communities like Grays Ferry in Philadelphia pushing back over environmental and health concerns.
Infrastructure strain. BloombergNEF projects U.S. data centers could consume about 18 Bcf/d of natural gas by 2035, nearly double the prior forecast, while hyperscaler AI data center spending may reach nearly $1.1 trillion by 2027. A Wharton analysis says earnings must grow productivity by a factor of 2.7 by 2030 to break even.
That's everything for this day – about a 5-minute read.