AI 'Scientist' Automates Research Process
A new AI system named Kosmos can autonomously form hypotheses, read over 1,500 research papers, and write 42,000 lines of code in just 12 hours to test its ideas. The system uses a structured world model to prevent common AI failures, showcasing a potential future for automated scientific discovery.
Kosmos was developed by Edison Scientific, the for-profit spin-off of FutureHouse, a non-profit research group funded primarily by former Google CEO Eric Schmidt. Collaborating scientists reported that a single 20-cycle run of Kosmos can replicate, on average, six months of their own research time, with independent verification finding nearly 80% of the AI's statements to be accurate. The system's core innovation is its "structured world model," which acts as a central information hub. This allows separate AI agents—one analyzing data and another searching scientific literature—to maintain coherence over hundreds of actions, a key failure point in previous agentic AI systems. This architecture enables the AI to build an internal representation of a scientific problem's structure and dynamics, similar to how world models are used to teach robots about physics. Among its initial discoveries, Kosmos identified a previously unreported linear relationship for improving perovskite solar cell efficiency and independently developed a method for prioritizing causal mechanisms from genetic data, identifying a protective pathway in Type 2 diabetes. In neuroscience, it analyzed transcriptomic data to infer that specific vulnerable neurons signal microglia to be consumed. This leap in automated research aligns with Turkey's strategic push into AI, underscored by the government's plan to create a $10 billion venture fund by 2030 to finance AI and data processing startups. The initiative aims to build out data centers and cloud infrastructure to support AI applications in key sectors like finance, healthcare, and industry. The Turkish AI startup scene is already gaining momentum, with approximately 1,200 active companies, 70% of which were founded after 2020. Despite a broader venture slowdown in 2025, AI startups saw increased funding, with notable rounds like mediatech platform Fal.ai's $49 million Series B and B2B intelligence startup HockeyStack's $20 million Series A. These developments are critical for Turkey's deeptech sector, where commercializing university research remains a key challenge. While deeptech constitutes less than 5% of the local startup ecosystem, incubators like Teknopark Istanbul's Cube Incubation and EU co-financed programs like Ege D-Tech are being established to provide the financial and institutional infrastructure needed to bring these complex ventures to market. In climatetech, AI-driven data analysis is crucial for industrial decarbonization and grid management as Turkey rapidly expands its renewable energy capacity. The country has already achieved nearly a third of its 2035 solar and wind power targets, with renewables supplying 56.7% of its electricity in 2025. This growth occurs within a challenging macroeconomic environment, with high inflation and interest rates impacting investor confidence. Venture funding in Turkey remains heavily concentrated in the seed stage, with a low conversion rate to Series A and beyond, prompting many later-stage companies to relocate abroad to access deeper capital pools.
