Every major decision rests on a forecast, and Radical Ventures portfolio company Mantic is building AI that turns those forecasts into something investors and operators can act on. In the latest Radical Talks episode, Radical Partners Rich Kotite and Aaron Rosenberg speak with Mantic co-founders Toby Shevlane and Ben Day about Mantic’ssummer’s system that outscored all 676 human participants, including professional forecasters, in this Metaculus Cup.
Mantic tests every change against questions with known outcomes, measuring in a morning what would take human forecasters a year to evaluate. Its customers include hedge funds and private equity investors who use Mantic’s forecasts in their decision-making systems.
Listen to the full conversation on Apple Podcasts, Spotify, or YouTube.
AI News This Week
- AI Disruptors 60: The Class of 2026 (TechCrunch) — The annual 2026 AI Disruptors 60 list, unveiled at TechCrunch Disrupt, recognizes four Radical Ventures portfolio companies. Etched, Cyera, and Periodic Labs were named among 15 Pioneers leading AI at scale. Etched designs transformer-specific chips, Cyera offers an AI-native data security platform, and Periodic Labs is building AI scientists for materials discovery and the autonomous laboratories they run in. Included in the 2026 list is Ricursive Intelligence, which develops reinforcement learning systems that automate chip design.
- Cohere Moves to Buy AI Cybersecurity Startup Aisle (Reuters) — Radical Ventures portfolio company Cohere has signed a letter of intent to acquire Aisle, a Prague-based startup that says its AI systems found over 500 previously unknown vulnerabilities in over 100 widely used open-source software codebases this past year, including many that Anthropic’s models did not detect. Cohere plans to pair Aisle’s autonomous security research with its foundation models, betting that the two can reinforce each other. Cohere co-founder Ivan Zhang explains that as AI grows more capable at both attack and defence, defenders hold the advantage because patching vulnerabilities is easier than discovering and exploiting them.
- Neocloud Prime Intellect Acquires Startup To Speed Up Inference (The Information) — Radical Ventures portfolio company Prime Intellect has acquired Moondream, a startup that builds small, open-weight vision models. Prime Intellect’s view is that a model’s value depends on how it is served, because inference determines both speed and cost. The deal is a bet that specialized models built alongside Prime Intellect’s serving infrastructure can outperform larger general-purpose models on specific tasks. In Prime Intellect’s internal testing, Moondream’s latest model reached 80% of a leading frontier model’s accuracy at locating objects in images, at roughly 1/1,000th of the cost.
- Promise Robotics strikes deal to supply largest US homebuilder DR Horton (BetaKit) — Homebuilders facing labour shortages and rising costs are turning to robotic factories to build more homes faster. Radical Ventures portfolio company Promise Robotics has signed a multi-year agreement with DR Horton, the largest US homebuilder by volume, to help build more than 1,250 homes around Phoenix. Promise’s platform uses AI to turn existing home designs into production plans and robotic instructions. The company’s off-the-shelf robots fitted with its own tooling assemble walls, floors, stairs, and roofs. Promise will begin manufacturing at a new Phoenix factory, its first outside Canada, in early 2027 and says it is contracted to build over 2,500 homes for several builders.
- Research: Harness Learning Enables Generalizable Test-Time Adaptation (John Hopkins/CMU) — An AI agent’s performance depends on both its model and its harness, the code that controls what the model sees and which tools it can use. To adapt harnesses to unfamiliar tasks, researchers trained a small proposer model with reinforcement learning to read execution reports and rewrite harness code. On reasoning tasks excluded from training, the trained 4-billion-parameter proposer nearly doubled the untrained version’s average score and slightly outperformed the 35-billion-parameter model that supplied its training examples. The authors see harness learning as a path toward agents that turn accumulated experience into improvements without updating model weights.
Subscribe to Radical Reads for our weekly insights from the frontier of AI
Radical Reads is edited by Ebin Tomy (Analyst, Radical Ventures)