AI & design briefing for Thursday, August 20, 2026

tldr

• Merck and Moderna report the first positive Phase 3 result for an AI-designed individualised mRNA cancer therapy; Generalist drops GEN-1.5 enabling robots to learn tasks from a single 3-second demo; and Anthropic surpasses OpenAI in quarterly revenue while OpenAI targets a 2027 IPO.
• Rich Sutton argues AI must learn continually from the world rather than static datasets; the FTC proposes treating secret personalised pricing as deceptive; Jessica Walsh examines whether design is dead in the age of AI on Adobe Design; and Flock Safety launches police AI for natural-language surveillance searches.
• Cursor launches Subscriptions for auto-fixing PR feedback; Google offers free Gemini Advanced for students; Meta AI for Mac gains screen sharing and cross-app dictation; and Poline, an esoteric colour palette generator, ships as a lightweight TypeScript library.

Top Stories

• AI-designed mRNA cancer therapy clears Phase 3 in landmark Merck-Moderna trial. Merck and Moderna reported the first positive Phase 3 result for an individualised mRNA cancer therapy. Moderna's AI algorithms analyse a patient's tumour and blood sequencing data to predict up to 34 neoantigens most likely to trigger an immune response, which are then encoded into a custom mRNA treatment paired with KEYTRUDA. The melanoma trial showed significant improvements in recurrence-free and distant-metastasis-free survival versus KEYTRUDA alone. ^[1]
• Generalist GEN-1.5 lets robots learn new tasks from a single 3-second physical demonstration. Generalist released GEN-1.5, a robot model that watches a 3–12-second physical demo and immediately attempts the task, no gradient updates required. Across 10 simple tasks, one demo achieved 59% average success; 10 weight updates on five minutes of data raised that to 83%. The model can copy human-hand demos, use simulated demos on real hardware, and improvise with unseen tools, with weight changes under 0.15%. ^[2]
• Anthropic passes OpenAI in quarterly revenue; OpenAI targets 2027 IPO. Anthropic reported $11.6B in Q2 revenue versus OpenAI's $6.7B and posted a small operating profit, while OpenAI CFO Sarah Friar told employees the company expects to go public in 2027 or sooner if growth keeps accelerating. ^{[3] [4]}
• Unitree Robotics surges 542% in Shanghai IPO debut. Chinese robotics company Unitree, known for its backflipping humanoid robots, jumped 542% in its Shanghai debut after an IPO that raised roughly $905M. ^[5]

Deeper Cuts

• Rich Sutton: AI's next leap is continual learning from the real world, not bigger datasets. In a video talk, AI pioneer Rich Sutton argued that the next major advance in intelligence will come from agents that keep learning from the world they operate in, rather than training on ever-larger static datasets. Generalist's GEN-1.5 release the next day made this thesis more concrete, though its one-shot mode is in-context learning, not persistent weight-based learning. ^[6]
• FTC proposes treating secret personalised pricing based on private consumer data as deceptive. The US Federal Trade Commission released an enforcement policy statement proposing that companies using private consumer data to set personalised prices without disclosure could be engaging in deceptive practices. ^[7]
• Flock Safety launches police AI for natural-language searches across surveillance data. Flock Safety built an AI system that lets law enforcement search movements, associates, arrest records, dispatch logs, and commercial identity data using natural language queries. ^[8]
• AI companies buying and destroying rare books for training data, investigation finds. A 404 Media investigation found AI companies purchasing rare books, scanning them for training data, and then shredding the originals; an AirTag traced one shipment to an Amazon-owned facility. ^[9]
• Chinese AI firms accessed advanced Nvidia chips through overseas cloud loophole. Chinese AI companies accessed advanced Nvidia compute through overseas cloud providers, exposing a remote-access loophole in US chip export controls. ^[10]
• Is design dead in the age of AI? Jessica Walsh weighs in. Jessica Walsh writes on Adobe Design that design is not dead but undergoing a fundamental revision in the age of AI, exploring how the craft, role, and value of designers are evolving. ^[11]
• Do you miss the old internet? pi.fyi offers an alternative. A nostalgic reflection on pre-algorithmic internet culture, before vibeless curation, tech bros, and influencers took over, a return to the web as it was. ^[12]
• Shading motion: detecting and stylising movement with WebGPU and TSL. A technical deep-dive on detecting, extracting, and stylising motion using WebGPU and TSL, covering motion masking, blob tracking, optical flow, and velocity-based motion blur. ^[13]

Tools and Launches

• Cursor Subscriptions auto-fix PR feedback. Cursor's cloud agents now subscribe to PRs they create, waking back up when CI checks fail or bot feedback arrives, making fixes without a fresh prompt. Set a goal like 'keep this PR merge-ready' and walk away. ^[14]
• Beautiful.ai: AI presentation tool with auto-reformatting slides. Beautiful.ai turns content into polished slides that automatically reformat as you edit. 14-day free trial, then $12/mo billed annually. ^[15]
• Berd: cross-model agent workspace for Goose, Claude Code, and Codex. Berd keeps projects, skills, credentials, and conversation history in one desktop home, reusable across multiple agent apps. Apache 2.0. Free on macOS, Windows, and Linux. ^[16]
• Ornith-1.5: MIT-licensed self-improving open models up to 397B parameters. Ornith released 9B, 35B, and 397B open models under the MIT license featuring self-improving capabilities for research and commercial use. ^[17]
• Soniox: voice cloning and expressive TTS in 60+ languages. Soniox offers voice cloning and generates expressive speech across 60+ languages for production voice applications. ^[18]
• Taku: one-click desktop mini-apps from reusable AI workflows. Taku turns reusable AI workflows into one-click desktop mini-apps that you can run, remix, and share without config files. Free to start. ^[19]
• Google offers free Gemini Advanced for college students. Eligible college students receive one year of a paid AI plan including study notebooks, visualisations, and Deep Research, free for 12 months. ^[20]
• Google Search generative UI with interactive visuals and simulations. Google Search's AI Mode can now turn complex questions into custom interactive visuals, tables, graphs, or simulations, rolling out free in Search. ^[21]
• Alexa+: cross-device AI assistant for browser, Echo, and Fire TV. Alexa+ lets you plan, draft, create images, and carry conversations across your browser, Echo, and Fire TV. Free to try; free with Prime. ^[22]
• Meta AI for Mac adds screen sharing and cross-app dictation. Meta's desktop AI assistant for Mac now includes screen sharing, cross-app dictation, ad analysis, and Google Workspace workflow integration. ^[23]
• Poline: esoteric colour palette generation as a lightweight TypeScript library. Poline is a dependency-free, fast JavaScript colour palette generator written in TypeScript, enabling esoteric and harmonious colour scheme creation. ^[24]
• How to choose a UI font: 10 alternatives to Inter. A practical guide to selecting UI typography based on cap-centred vertical metrics and tall x-height, with ten alternatives to the ubiquitous Inter typeface. ^[25]

Prompt of the day