In this week’s edition of Radical Reads, we feature a conversation between Mehdi Ghissassi, Chief Product and Technology Officer at ai71 and former Head of Product at Google DeepMind, and Radical partner Aaron Rosenberg on turning frontier research into products people use.
At DeepMind, Mehdi helped bring research like AlphaFold and WaveNet out of the lab and into the real world. In this conversation, Mehdi and Aaron explore why customers pay for outcomes over raw model capability, why chasing too many use cases early can slow a company down, and where smaller companies can compete with the large frontier labs.
Listen to the full conversation on Apple Podcasts, Spotify, or YouTube.
AI News This Week
- Ottawa Buys $1.6-billion Stake in Cohere as AI Firm’s Valuation Soars to $35.7-billion (Globe and Mail) — Radical Ventures portfolio company Cohere is on the cusp of closing a $7 billion financing round at a $35.7 billion valuation, the largest raise by a Canadian startup in history. The round is led by Germany’s Schwarz Group. The Canadian government is contributing $1.6 billion for an equity stake, an unprecedented direct investment in a technology company. In return, Cohere has committed to spend $1.7 billion on domestic compute infrastructure by 2030 and to keep its headquarters in Canada. Annualized revenue was growing 165% as of September, and the company’s customer base increased 20-fold to 300 clients over the past year.
- Inside Giant Homebuilder D.R. Horton’s Bet on Robotic Homebuilding (FC) — Radical Ventures portfolio company Promise Robotics has signed a multi-year agreement with D.R. Horton, America’s largest homebuilder, to help build more than 1,250 homes in metro Phoenix. Under its factory-as-a-service model, Promise Robotics owns and runs the factories and technology while builders purchase the output. Its AI turns a builder’s existing plans into machine instructions, and robots produce framing components that are then installed on site.
- Did the Hardest Problem in A.I. Just Get Much Easier? (NYT) — Interpretability, the science of explaining why AI models behave as they do, is key to trusting AI in sensitive settings and grows harder as models scale. In a 26,000-word essay, former OpenAI researcher Nick Cammarata theorizes that as networks grow larger, individual neurons increasingly line up with specific concepts such as shapes or sounds. Looking again at Google’s Inception image model, he says he can explain the role of almost every neuron.
- Jane Street and Rivals Are Now Fighting for AI Power as Well as Pace (Bloomberg) — High-frequency trading firms that built their businesses on speed are now competing for compute. Firms are building or leasing their own data centers as machine learning becomes central to predicting market behaviour and attracting top research talent. Jane Street recently signed a $13 billion agreement with Radical Ventures portfolio company Crusoe and led a funding round in AI chipmaker Etched, part of a broader push by trading firms to back the infrastructure and chip startups that could give them an edge.
- Research: Cache Engineering for Agents (Tsinghua/CUHK) — As AI agents take on tasks spanning hundreds of steps, storing and rereading their accumulated working memory, known as the KV cache, becomes the largest share of inference cost. The common fix of trimming or summarizing old history forces expensive recomputation. Researchers propose cache engineering, which lets agents drop selected cached content, such as older tool outputs, while reusing everything they keep.
Subscribe to Radical Reads for our weekly insights from the frontier of AI
Radical Reads is edited by Ebin Tomy (Analyst, Radical Ventures)