🧠 Lab Coats, Open Models, and Your Inbox

NMR spectroscopy is the standard tool chemists use to figure out the structure of a molecule — think of it as an MRI for chemistry. Interpreting NMR data correctly is painstaking, technical work, and there's specialized software (like ChemDraw and MestReNova) that's been built specifically for this job.

Anthropic published a research report showing that Claude Opus 4.7 can perform NMR spectroscopy tasks at a level that matches, and in some cases exceeds, those dedicated tools.

The study tested Opus 4.7 across 20 compounds sourced from recent synthetic chemistry preprints, evaluating both forward prediction (simulating what a spectrum should look like for a given structure) and inverse structure elucidation (working backward from spectral data to figure out the molecule). On hydrogen NMR shifts, Opus 4.7 posted the lowest average error at ±0.079 ppm.

Why it matters: this isn't a niche research flex. Chemistry is one of the slowest parts of drug discovery and materials science. A general model that can reliably do this work means researchers spend less time translating data and more time running experiments.

Most AI models are built for single-turn conversations — you ask, they answer, done. But the next wave of AI is agentic: models that run for extended periods, use tools, maintain context across dozens of steps, and coordinate with other models to complete complex tasks. That's a fundamentally different design challenge.

NVIDIA released Nemotron 3 Ultra, a 550B-parameter Mixture-of-Experts model with 55B active parameters, optimized for orchestrating complex, long-running agent workflows by combining frontier reasoning and high throughput with domain adaptability.

The model supports three reasoning modes — reasoning-off, regular, and medium-effort — and also accepts inference-time budget control. The medium-effort mode uses about 2.5x fewer tokens than regular mode, with roughly a 7% drop in accuracy. That efficiency lever matters a lot when you're running agents at scale.

The model weights, over 10 trillion tokens of datasets, and training recipes are all being released openly.

The workflow most people have been running: ask ChatGPT to draft an email, copy the output, open Gmail or Outlook, paste, tweak, send. It's a small friction — but it adds up.

OpenAI has rolled out the ability to send emails directly from writing blocks in ChatGPT on the web, keeping the whole process inside a single conversation from start to finish. Writing blocks are a dedicated editing interface inside ChatGPT that presents drafts as a clean, formatted document rather than a plain chat reply.

Users connect their email account via settings, and can then send drafts to recipients with a single click within the writing blocks section. The feature is rolling out gradually to web-based ChatGPT users. It's a small upgrade on paper, but it signals something bigger — AI tools are moving from assistants that suggest to assistants that act.

Paper is a design tool that works with real, editable SVG files — not rasterized images. SVG (Scalable Vector Graphics) is a file format that keeps shapes and paths mathematically defined, meaning they scale to any size without losing quality. Designers use SVGs for logos, icons, and UI elements.

The Pen Tool just went live in Paper, letting you draw shapes, generate SVGs from prompts, or vectorize existing images — and then edit the actual points and paths directly. Because Paper exports real SVGs, what you see is exactly what you get. Best for designers and developers who want a faster SVG workflow without the complexity of Figma or Illustrator.

Text-to-speech tools convert written text into spoken audio — used everywhere from voice assistants to audiobook generation. Most of them still sound robotic or flat when handling emotional nuance.

Miso One is an 8-billion-parameter text-to-speech model built specifically for expressive, human-sounding speech. It responds with just 110ms of latency (faster than a human speaking naturally) and is built to match the emotional tone of what it's reading. The model weights have been open-sourced, with API access coming soon. Best for developers building voice products, content creators doing AI voiceover work, or anyone who's given up on TTS sounding human.

Gotcha is an app built using Claude Code (Anthropic's agentic coding tool) that turns your phone's camera into a creature-catching device — except instead of fictional Pokémon, you're logging real animals you encounter in the world.

Here's how it works:

It's Pokémon Go, but the world is the game. And it was built by one person using AI-assisted coding.