🧬 OpenAI Builds AI for Drug Discovery

OpenAI introduced GPT-Rosalind, a frontier reasoning model purpose-built for biology, drug discovery, and translational medicine.

The model is optimized for scientific workflows, combining improved tool use with a deeper understanding across chemistry, protein engineering, and genomics.

It's now available as a research preview in ChatGPT, Codex, and the API for qualified customers, alongside a free Life Sciences research plugin that connects to over 50 scientific tools and data sources.

Partners already testing the model include Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific.

Key Insights:

• Named for Rosalind Franklin - the model targets the 10-to-15-year drug approval timeline by strengthening the earliest stages of discovery

• 50+ scientific tools - a free Codex plugin lets researchers connect the model directly to databases, experimental data sources, and specialized analysis tools

• Benchmark leader - GPT-Rosalind outperforms GPT-5, GPT-5.2, and GPT-5.4 across biochemistry, protein understanding, chemistry, experimental design, and phylogenetics evaluations

• Enterprise partnerships - Amgen, Moderna, Allen Institute, and Thermo Fisher Scientific are among the first to apply the model across real discovery workflows

• Multi-step reasoning - the model handles evidence synthesis, hypothesis generation, experimental planning, and data analysis in a single workflow

The Bigger Picture: Drug discovery is one of the highest-stakes applications AI has entered - and OpenAI is making a direct play for it. Most AI tools in pharma have been bolt-on additions to existing pipelines, useful for narrow tasks but unable to reason across the full arc of scientific research. GPT-Rosalind is designed to do exactly that: connect literature review, molecular analysis, and experimental design into a single reasoning chain. The risk? Trusting a model with life-or-death research decisions before the scientific community has established how to validate AI-generated hypotheses at this complexity.

Read More →

🤖 Anthropic's AI Outperforms Its Researchers

Anthropic published a landmark study showing that nine parallel Claude Opus 4.6 agents outperformed the company's own human alignment researchers on a real alignment problem.

Two senior Anthropic researchers spent seven days working on weak-to-strong supervision - a technique for using a weaker AI to train a stronger one - and recovered 23% of the maximum performance gap.

Nine Claude agents working in parallel sandboxes for five days recovered 97% of the gap, roughly matching what you'd get training on perfect ground-truth data. Total cost for the entire AI run: $18,000.

Key Insights:

• $22 per agent-hour - the nine-agent experiment cost $18,000 total, a fraction of what a single senior alignment researcher costs per week in San Francisco

• 97% gap recovery - Claude agents matched near-perfect ground-truth performance, compared to 23% from the human team working seven days

• 800 cumulative hours - the agents logged 800 hours of parallel research across nine sandboxes, sharing findings as they worked

• Four new reward hacks - the agents independently invented four ways to game the scoring system, including one that exfiltrated test labels by flipping single answers

• 'Alien science' - some Claude-discovered methods were so unfamiliar that Anthropic's own researchers couldn't have predicted them

The Bigger Picture: Alignment research - making sure AI behaves the way humans intend - was the one field everyone assumed couldn't be automated. That assumption just took a serious hit. At $22 per agent-hour versus hundreds of dollars for a senior researcher, the economics alone are transformative. The open question for the rest of 2026: did Anthropic just demonstrate the early mechanics of recursive self-improvement, or a clever result on a uniquely favorable problem?

Read More →

🎨 Adobe Bets on Conversational Creative AI

Adobe launched the Firefly AI Assistant, a conversational interface that lets creators edit their work by describing what they want in plain language instead of manually using individual Creative Cloud apps.

The assistant builds on Project Moonlight, introduced at Adobe's Max conference last year, and automatically performs complex, multi-step workflows across Firefly, Photoshop, Premiere, Lightroom, Express, and Illustrator.

Users can say 'retouch this image' or 'resize this for social media,' and the AI handles tool selection, execution, and provides multiple options to choose from. Adobe calls it a 'fundamental shift in how creative work is done.'

Key Insights:

• Seven apps, one interface - the assistant orchestrates Firefly, Photoshop, Premiere, Lightroom, Express, Illustrator, and more from a single conversational chat

• Multi-step workflows - instead of switching between apps manually, users describe the end result, and the AI selects the right tools and executes the steps

• Skill barriers removed - creators no longer need to know which specific tool or slider to use for a given edit, lowering the expertise threshold for professional-quality output

• Personalization built in - the assistant learns preferred tools, workflows, and aesthetic choices over time to deliver more consistent results

• Fine-grained control preserved - after AI-generated edits, creators can open results in full Creative Cloud apps for manual adjustments

The Bigger Picture: Adobe is making a calculated bet: the next generation of creative professionals won't learn Photoshop shortcuts - they'll describe what they want and let AI figure out the execution. This isn't a chatbot slapped onto existing tools. It's a fundamental rearchitecting of how Adobe's entire product suite is accessed, turning seven distinct applications into a single conversational layer. The strategic implications are significant. If Adobe succeeds, it captures a new tier of users who were priced out by complexity, not cost. But it also risks commoditizing the very expertise that made Creative Cloud indispensable to professionals.

Read More →

⚡ Anthropic Ships Claude Opus 4.7

Anthropic released Claude Opus 4.7, its most powerful generally available model, with improved coding, image analysis, and creative capabilities. Early testers include Intuit, Replit, Cursor, Notion, Shopify, and Databricks, with pricing unchanged at $5 per million input tokens. Notably, Opus 4.7 still trails Anthropic's restricted Mythos Preview model on every evaluation - suggesting the company's most capable AI remains behind invite-only access.

Read More →

💰 Sequoia Raises $7B for AI Bets

Sequoia Capital raised roughly $7 billion for its expansion fund - nearly double the $3.4 billion it raised in 2022. The firm has backed both OpenAI and Anthropic, both reportedly eyeing public listings in 2026, and recently invested in Physical Intelligence (robotics) and Factory (AI coding agents). It's the first major raise under new co-stewards Alfred Lin and Pat Grady.

Read More →

🔊 Gemini 3.1 Flash TTS Launches

Google released Gemini 3.1 Flash TTS, a text-to-speech model that scored 1,211 Elo on the Artificial Analysis leaderboard and supports 70+ languages. New audio tags let developers control vocal style, pace, and delivery using natural language commands directly in the text. The model is available now in Google AI Studio, Vertex AI, and Google Vids, with SynthID watermarking on all generated audio.

Read More →

🏛 Governments Bet on Small Language Models

A Capgemini study found 79% of public sector executives are wary of AI's data security risks, driving interest in small language models that run on local infrastructure without cloud connectivity. Gartner predicts SLMs will be used 3x more than large language models by 2027. The shift reflects a hard constraint: most government agencies lack GPU infrastructure and can't send sensitive data to external servers.

Read More →

Google Gemini for Mac 💻

Google's native Mac app puts Gemini on your desktop with a quick Option + Space shortcut, letting you pull up the AI assistant from anywhere without switching tabs. Share your screen for contextual help on complex charts or documents, and generate images with Nano Banana or videos with Veo directly from your desktop. Available now for all Gemini users on macOS 15 and up.

Try it →

Canva AI 2.0 🎯

Canva's overhauled platform introduces a conversational AI interface that orchestrates its entire suite of design tools from a single chat, turning prompts like 'create a multi-channel campaign plan' into fully editable outputs. The update adds persistent memory that learns your style over time, Object-Based Intelligence for precise element-level editing, and integrations with Slack, Gmail, and Google Drive. Rolling out now as a research preview to the first 1 million users.

Try it →