Perplexity Launches Self-Improving Context Graph, Midjourney Announces Medical Division and Hardware, OpenAI Advances in Scientific Research

Key Takeaways

1. Perplexity's Brain context graph enables self-improving, stateful AI agents by capturing and refreshing context across tools, available to all Max subscribers.
2. Midjourney expands into hardware and medical AI, with a live reveal tomorrow and a new medical research division, signaling major platform ambitions.
3. OpenAI demonstrates real-world scientific impact through drug chemistry, rare pediatric diagnosis, and a new life sciences benchmark (LifeSciBench).
4. Anthropic's Opus 4.7 achieves 20x speedup over human teams in coding robotics tasks, though the robot dog still fails to retrieve a beach ball.
5. Open-source AI momentum surges with GLM-5.2 free access, Gemma 4 reaching 255 tok/s in browser, and new multilingual models from Hugging Face ecosystem.
6. Ethan Mollick warns that big company AI strategies are outdated (pre-agent era), and significant bets on exponential growth of top labs may be at risk.

1. Major Product Launches and Platform Expansions

1. Perplexity launched Brain, a self-improving context graph that automatically updates at night and feeds fresh context to every task on Computer, giving it state and self-improvement capabilities — now available to all Perplexity Max subscribers. — via 1 2
2. Runway introduced Recipes via API, enabling one-call integration of production-grade generative media features without building or maintaining custom workflows. — via 1
3. xAI made Grok Build available on DigitalOcean Marketplace (one-click VM deployment) and integrated Grok into Databricks Agent Bricks for enterprise AI agents. — via 1 2
4. Midjourney announced its first hardware project (live stream June 17, 6 PM PT), a new Medical division, and released a big-batch draft mode for V8.1 (24 low-res images at half price) plus a technical overview of its Scanner. — via 1 2 3 4

2. Key People Moves and Research Directions

1. Sam Altman welcomed Noam Shazeer to OpenAI, calling him one of the people he most wanted to work with since OpenAI's founding — a decade-long wait. — via 1 2
2. Yann LeCun joined AmiLabs as Paris Research Director, focusing on world models — a long-term direction aimed at grounding AI in physical reality, which he argues is a more promising path than language-only generative AI; he also predicted that in a decade everyone will have a team of virtual assistants. — via 1 2 3
3. Anthropic released the second phase of Project Fetch, where Opus 4.7 alone completed programming tasks ~20x faster than the best human team (assisted by Opus 4.1) from last year, though the robot dog still failed to retrieve a beach ball. — via 1

3. OpenAI's Scientific and Medical Breakthroughs

1. OpenAI's GPT-5.4 powered a drug chemistry project (via Maria AI and dedicated lab) that discovered an unexpected method improving a widely used reaction, from literature review to experimental validation. — via 1
2. OpenAI launched LifeSciBench, a benchmark with 750 expert-crafted tasks across seven biology workflows, developed with 173 biotech/pharma scientists, to measure AI support in life sciences. — via 1
3. In collaboration with Boston Children's Hospital and Harvard, o3 Deep Research helped clinicians re-examine unresolved rare pediatric cases, providing answers for families who had waited years — published in NEJM AI. — via 1

4. Open-Source AI Momentum and Industry Insights

1. Llama.cpp received a new brand and website, with Hugging Face emphasizing that open source must win. — via 1
2. Rising costs are driving increased interest in open-source AI, with Chinese companies leading in this space. — via 1
3. Gemma 4 achieves 255 tokens/second in a browser; agentic kernel optimization is called the future of on-device inference. — via 1 2
4. GLM-5.2 is available for free, with the claim that open source has caught up. — via 1
5. MOSS-TTS v1.5 supports 30+ languages at 48kHz with voice cloning. — via 1
6. LFM2.5 multilingual retrieval model supports 11 languages with 1.5ms end-to-end latency. — via 1
7. Ethan Mollick observed that big company AI strategies were set before the agent revolution (late 2025) and are now outdated; significant capital bets on continued exponential growth of top three labs may be at risk, and that Google no longer has a public frontier model (Gemini 3.1 Pro lagging). — via 1 2 3
8. Deedy noted that Meta is turning 30-50% of core SWEs into data labelers for AI-generated GitHub repos, sarcastically calling this a training data generation exercise. — via 1