The State of AI Assistants
Research Rundown #149, plus new memos on Stripe, Astranis, and more.
Research Rundown
From Iron Man’s J.A.R.V.I.S. to Samantha from Her, AI assistants have long captured the public's imagination. But despite having dreamed the dream of what an advanced fully personalized AI assistant could be capable of, today’s assistants are lackluster in comparison. In our latest deep dive, we unpack the dramatic limitations of the first generation of NLP-based AI assistants like Siri and Alexa, and the new generation of generative-based AI assistants that are hoping to automate some of the more mundane aspects of our lives.
Despite some meaningful new capabilities, the reality is that the gap between the science fiction portrayal of AI assistants and reality is still quite wide. This gap is not due to a failure of imagination but a reminder that reliable intelligence requires foundations that are still being built.
Read our full deep dive here.
How did Stripe go from a simple payments tool to the invisible engine behind billions in global commerce? Read our memo here for everything you need to know.
Astranis is betting small satellites can solve one of the world’s biggest problems—will it work? We break it down in our memo here.
Behind the scenes, Peregrine is turning messy police records into the real-time insights powering modern public safety. To find out more, read our memo here.
OpenAI CFO Sarah Friar said the company may eventually rent out its AI data centers, echoing Amazon’s cloud model, though the focus now is on meeting its own needs. CEO Sam Altman expects trillions in future data center spend, financed through debt, private equity, and new instruments.
Google measured the environmental footprint of AI inference across its full serving stack and found that a median Gemini Apps prompt in May 2025 used 0.24 Wh, 0.03 g CO₂e, and 0.26 mL water. Efficiency gains, clean energy, and hardware advances cut energy use 33× and emissions 44× over the past year, showing optimization can sharply reduce AI’s impact.
Meanwhile, Louisiana regulators approved Entergy’s plan to build three natural gas plants and transmission lines to power Meta’s 4-million–square-foot Hyperion data center, which could draw 5 gigawatts. Meta pledged to cover costs and offset use with renewables, while Entergy will add 1.5 gigawatts of solar. The deal was expedited to keep Meta in-state, but critics warn it will raise bills and harm the environment.
Meta Superintelligence Labs chief Alexandr Wang announced a major reorg to speed progress toward “superintelligence,” creating four teams: TBD Lab for large-scale training, FAIR as an innovation engine, Products & Applied Research under Nat Friedman, and Infrastructure under Aparna Ramani. The shift dissolves AGI Foundations and puts most leaders reporting directly to Wang.
Perplexity CEO Aravind Srinivas said AI agents will take over tedious online tasks, letting people browse mainly for curiosity and discovery. He sees curiosity as the key human trait in the AI era and is betting on Comet, a browser that embeds AI for faster, context-aware actions. Long-term, he views the browser as a path to an AI-native operating system, positioning Perplexity against Google, Apple, and OpenAI.
Databricks is raising funds at a $100 billion valuation, up 61% from December, with Thrive, Insight, WCM, and Andreessen Horowitz investing. The AI-driven growth supports new SAP and Palantir partnerships, AI-focused databases, and 3,000 new hires. CEO Ali Ghodsi said demand lets Databricks delay an IPO and aim for trillion-dollar scale.
Researchers from Apple and Stanford introduced a foundation model trained on over 2.5 billion hours of wearable behavioral data from 162K participants, showing that it outperforms traditional biosignal models across 57 health tasks, including sleep, injury, infection, and pregnancy.
Ro announced Serena Williams as its newest ambassador, with CEO Z Reitano highlighting that her weight loss journey, despite training and diet, required additional support from GLP-1 treatment. By sharing her story, Williams aims to destigmatize medication use in health and inspire others navigating similar challenges.
In a new essay from a16z, Sarah Wang and Martin Casado argue early low margins in AI apps are temporary. Tiered pricing, model routing, and enterprise contracts improve economics, while competition and declining inference costs ease pressure. Leading apps differentiate with features, data, and fine-tuned models, driving long-term margin expansion.
AI model costs have stalled, pressuring developers but boosting Microsoft, OpenAI, Google, and Anthropic. Intuit’s Azure spend will hit $30M this year, while Microsoft’s Azure revenue rose 39% on AI demand. Startups like Cursor and Replit raised prices as GPT-5 and Anthropic models stay costly, leaving buyers burdened despite OpenAI offering cheaper alternatives like o3.
An MIT report found 95% of enterprise GenAI pilots failed due to poor integration and workflow fit, while narrow use cases and vendor partnerships had higher success. The results shook investor confidence in the AI boom, though firms like IBM and Accenture stand to benefit.
At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!
Seems to me that Axon is a serious competitor for Perigrene.