Google DeepMind’s AI Co-Scientist now plans experiments, operates laboratory equipment and writes scientific papers. The Gemini-based multi-agent system has expanded from generating hypotheses into a research workflow integrated with the lab, according to The Decoder.
The system’s work spans three disciplines, including materials synthesis and the autonomous development of a medical AI architecture. That gives Co-Scientist a broader role than a research assistant that only proposes ideas: it now handles parts of experiment execution and research communication.
The report provides no performance benchmarks, deployment timeline or evidence that Co-Scientist can independently complete an end-to-end research program. The claim here is scope, not a demonstrated replacement for laboratory scientists.
Why it matters: Google DeepMind’s Co-Scientist now connects hypothesis generation, experiment planning, lab-equipment operation and paper writing across three disciplines, but missing benchmarks leave researchers without evidence of a practical advantage.
Lambda has raised $1 billion in private debt to buy Nvidia AI chips and lease them to Microsoft, according to TechCrunch AI. The neocloud provider is using borrowed capital to buy more chips for its leasing business.
The financing is the latest in a string of loans for Lambda. TechCrunch AI said the borrowing underscores the high cost of the AI boom.
Why it matters: For Lambda, the deal adds another private loan to the financing behind Nvidia-chip leases for Microsoft.
About 70% of new AI data-center watts are going into the United States, while China is getting under 10%. That allocation leaves the U.S. with more than 10 times China's AI compute.
In 2022, the U.S. accounted for roughly 45% to 50% of the world's new compute, compared with China's 30% to 35%. Export controls and a huge U.S. buildout flipped that balance, while U.S. AI chips now deliver 2.3 to 2.7 times better compute per watt.
Why it matters: For U.S. and Chinese AI developers, the 70%-versus-under-10% allocation is concentrating new compute in America; export controls and a huge U.S. buildout drove the reversal.
One person using a chatbot burns tokens only while typing; one person running 3–20 background bots burns tokens all night, according to Next Big Future AI.
Cheap inference makes that overnight workload more affordable. GLM-5.3-Flash is listed at $0.15/$0.50 per million tokens, roughly 10x cheaper than a frontier model. Next Big Future AI calls this month's development the start of a “grokbot boom.”
Why it matters: For a single user, GLM-5.3-Flash's listed rate—roughly 10x below a frontier model—makes running 3–20 background bots overnight more affordable.
Wired AI reports on a recent paper arguing that AI is often better at doctoring than human doctors. The article offers no details about the paper beyond that claim.
Its reported reaction is blunt: doctors are not thrilled. The paper also leaves physicians with a direct question—what remains for them if AI can do the work better?
Why it matters: The paper challenges doctors’ professional role by asking what remains for physicians if AI can perform doctoring better than they can.
Decathlon forecasts weekly demand for tens of thousands of products across multiple continents. The retailer deployed Chronos-2 on AWS for the workload, according to the AWS ML Blog.
AWS says the deployment improved forecast accuracy by 11–15 percentage points, cut operational complexity, and ran weekly inference for about $0.03 on CPU-only instances.
Why it matters: For Decathlon’s demand-planning teams, a reported $0.03 weekly inference cost and 11–15-point accuracy improvement apply to forecasts covering tens of thousands of products across multiple continents.