Julian

Jancis Robinson’s latest 2016 Bordeaux tasting (10-year retrospective) shows the vintage is aging beautifully. The 2016s tasted at 10 years old are looking very snazzy—left-bank and top right-bank wines particularly strong, though some mid-range St-Émilion producers still making wine to “outdated and overworked formulas.” Lesser right-bank wines already drinkable with pleasure. Full tasting notes across right-bank reds, sweet whites, and left-bank reds available. Via Jancis Robinson. Read more

Jason Wilson explores why US wine prices jumped 11% in 2025. The real story: it’s not inflation—it’s that no one is buying cheap wine anymore. Shipments under dropped 67% over 5 years. Meanwhile, younger drinkers (the largest wine cohort now) spend more per bottle than boomers, with the sweet spot at –. The industry’s problem: it’s still trying to push cheap bulk wine instead of quality and experience. Why it matters for Julian: clear market signal that quality-focused wine is where the growth is. Via Everyday Drinking. Read more

Google DeepMind just launched Nano Banana 2, combining the advanced capabilities of their Pro model with the speed of Gemini Flash. The new image model offers world knowledge, subject consistency (up to 5 characters/14 objects), precision text rendering, and 4K resolution output—all at flash speeds, making rapid iteration possible. Rolling out across Gemini, Search, Google Ads, and Vertex AI. Why it matters: Makes production-ready image generation accessible at speed. Via The Rundown AI. Read more

Hilary Gridley argues that rushing to hire outside consultants to drive AI transformation signals to your team: “I am about to fire you.” Better approach: invest in upskilling your domain experts. They need time to experiment, clarity on priorities, ambitious goals, visible rewards for wins, and a culture of learning. The key insight: successful AI transformation requires your team’s brains to rewire, not just their tools to change. Via Hilary’s Substack. Read more

Premiere Napa Valley’s average bottle price dropped ~39% from 2023, and ,000+ lots disappeared. Cabernet Sauvignon still dominates volume, but Cabernet Franc averaged higher prices ( vs ), hinting at shifting demand. The downturn coincides with closures and vineyard sales in Napa, including Gallo’s Ranch House shutdown and Trinchero listing vineyards. Via Wine-Searcher. Read more

After the Supreme Court struck down most tariffs, Trump invoked Section 122 to impose a 15% tariff for up to 150 days. The wine industry faces paralysis as lawsuits loom and timelines are unclear; uncertainty is freezing import/export planning. The piece notes even US producers get hit because key winery equipment is imported. Via Wine-Searcher. Read more

Notes from the Pragmatic Summit and a 50-person workshop: 92% of devs now use AI coding tools monthly, and healthy orgs see ~50% fewer incidents while unhealthy ones see 2x more. The article covers what AI-native teams look like, a quiet crisis for mid-level engineers, why refactoring still matters, and a possible resurgence of XP practices. Via The Pragmatic Engineer. Read more

New paper argues safety testing treats jailbreaks as isolated bugs, but failures form large behavioral regions that persist across paraphrases. The authors propose mapping these “failure basins” using MAP-Elites to chart where models fail, how big those regions are, and where refusal flips to compliance. That shifts evals from incident counting to systems-level mapping. Via Agentic AI. Read more

OWASP’s new AI Vulnerability Scoring System (AIVSS) extends CVSS with agent-specific amplification factors like autonomy, tool access, and persistent memory, producing a 0–10 risk score across 10 agent security categories. It pairs with AIUC-1 controls so high-risk findings map directly to concrete mitigations and certification steps. Useful for quantifying agent risk in a board-friendly way. Via Agentic AI. Read more

Hilary Gridley makes a sharp case that AI shouldn’t help you do more — it should help you figure out which 2-3 things are actually worth doing. Her setup: Wispr Flow on the iPhone lock screen for instant idea capture, a Claude.md prompt that auto-organizes ideas, and a rapid validation workflow that tests many ideas quickly before committing. The goal is having 10x more ideas without splitting focus across 30 projects at 3% each. Via Hils (Substack). Read more

OpenAI and Jony Ive’s first physical product will reportedly be a $200-300 smart speaker with a built-in camera and facial recognition for purchases — aiming to ship by early 2027. The 200+ person team (formed after OAI acquired Ive’s Io Products for $6.5B) is building a device that observes surroundings and nudges users toward actions. Smart glasses are also planned but not until 2028. With Amazon’s Alexa+ and Apple AI devices already ramping, this is OpenAI’s first and very important swing at defining the AI hardware category. Via The Rundown AI. Read more

A practical engineering guide breaking down how Claude Code’s extension ecosystem is layered: Skills own reasoning orchestration, Tools provide scoped execution, Subagents handle parallelism, and Hooks enable event-driven actions. Includes YAML schemas for packaging skills, security constraints, and integration patterns for production agentic workflows. Useful context if you’re building on top of Claude Code or OpenClaw. Via Agentic AI (Ken Huang). Read more

Monarch Tractor, the autonomous electric tractor startup founded by Carlo Mondavi (Robert Mondavi’s grandson), has shut down. The company built battery-powered, self-driving tractors designed specifically for vineyard and farm use — a genuinely interesting bet on bringing autonomous tech to viticulture. Its closure raises questions about the viability of the ag-automation category in wine. Via JancisRobinson.com Wine News in 5. Read more

The Supreme Court ruled 6–3 today that the Trump administration’s reciprocal tariffs imposed under IEEPA were unconstitutional. The case was brought by wine importer VOS Selections. The ruling could lower import costs for European wines in the US — good news for anyone buying Italian, Spanish, or French bottles — though Trump has called the ruling a ‘disgrace’ and claims he has a backup plan. Worth watching how this plays out for pricing. Via JancisRobinson.com. Read more

Accenture is now monitoring AI tool logins weekly for associate directors aiming for leadership promotions, making AI usage a “visible input” to reviews. The target: senior employees, who resist AI adoption far more than junior staff. Of 780K employees, 550K have gone through AI training, but insiders reportedly call the internal tools “broken slop generators.” CEO Julie Sweet has previously said the firm would “exit” staff who don''’t reskill for AI. Via Financial Times / The Rundown AI. Read more

Google released Gemini 3.1 Pro, scoring 77.1% on the ARC-AGI-2 reasoning benchmark — more than double its predecessor’s 31.1%, and ahead of Claude Opus 4.6 (68.8%) and GPT-5.2 (52.9%). It also tops benchmarks for science, competitive coding, MCP use, and agentic search. Pricing is identical to 3 Pro with the same 1M token context window, available now via Gemini API, Vertex AI, the Gemini app, and NotebookLM. Via The Rundown AI. Read more

At India’s AI Impact Summit, PM Modi tried to link arms with tech leaders for a group photo — OpenAI’s Sam Altman and Anthropic’s Dario Amodei reportedly raised fists instead of joining hands. Altman said he was “confused.” The moment caps weeks of OpenAI/Anthropic friction, including Anthropic’s Super Bowl ad mocking ChatGPT ads, which Altman called “clearly dishonest.” A meme, but also a snapshot of how far from collaborative the top AI labs really are. Via The Rundown AI. Read more

The classic OSI model ends at Layer 7 (the application layer), where software executes deterministic, predictable operations. AI researcher Ken Huang argues agentic AI crosses a “Deterministic Boundary” that warrants its own conceptual Layer 8 — because unlike traditional apps, agents don’t just respond to inputs, they reason, plan, and act with autonomy. It’s not just another app running on Layer 7; it’s qualitatively different from anything we’ve built before. Via Agentic AI newsletter. Read more

Tavus introduced Phoenix-4, a real-time human rendering model that generates every pixel of an AI avatar’s face from scratch at 40 FPS — with 10+ emotional states, active listening behaviors, and millisecond-level expression transitions. Unlike older models that react to sound alone, Phoenix-4 understands conversational context and shifts expressions mid-conversation the way humans do. Tavus is pitching it for healthcare, education, and sales. The obvious flip side: this tech massively raises the ceiling on AI-generated deception in video calls. Via The Rundown AI. Read more

Google rolled out Lyria 3 in the Gemini app — its new AI music generation model that turns a text prompt or photo into a custom 30-second track, complete with auto-generated lyrics and cover art. Unlike dedicated tools like Suno or Udio, this puts AI music a single prompt away for millions of Gemini users who’ve never heard of those apps. Every track gets tagged with SynthID watermarking, and YouTube creators get access via Dream Track for Shorts. The model has been in development at DeepMind since 2023 — this is its first mainstream consumer release. Via The Rundown AI. Read more