Science+10Robots+5Aug 19

Image: The Threshold Report/GPT Image 2
Vivodyne opens autonomous lab for human-tissue drug tests
Vivodyne has opened an autonomous facility intended to help drug developers test candidate treatments on human tissue before committing to clinical trials. Its modular robotic HIVE labs grow tissues, apply doses, and monitor experiments across 20 tissue types. The company says the site produces causal data, meaning evidence about how biological states change after an intervention, for training models on ongoing experiments.
Audio+2Text+0Aug 19

Image: The Threshold Report/GPT Image 2
Calendly launches AI meeting recorder and note-taker
Teams that schedule through Calendly can keep call records and follow-up work in the same service. Its AI note-taker joins calls, records audio and video, transcribes the conversation, and produces summaries, action items, and draft follow-up emails. A chat alert tells participants it is present, and the tool leaves when asked.
Robots+2Aug 19

Image: The Threshold Report/GPT Image 2
Serve expands robot delivery through Grubhub and DoorDash
Restaurant customers can order sidewalk-robot delivery through services they already use in more locations. Serve is available through over 100 participating Grubhub merchants in Chicago and nearly 200 in Los Angeles, and Alexandria has joined that network. Its DoorDash service has also completed a first month of deliveries in Washington, D.C., and San Jose.
Text+2Aug 19

Image: openai.com
Replit opens free GPT-5.6 Luna software-planning mode
People without coding experience can explore and shape software ideas in the same Replit workspace used to turn them into applications. Free Mode uses GPT-5.6 Luna for planning, feedback, analysis, and other project work without counting against usage. Harder reasoning can move to GPT-5.6 Sol while keeping the project's context.
Compute & Infra+0Aug 19
Liquid AI releases smaller LFM2.5 files for local use
Smaller local models can run with less memory on modest hardware, making on-device assistants and tools more practical. Liquid AI released four 4-bit LFM2.5 GGUF model files (a format used by local model software) covering its 230M, 350M, 1.2B Instruct, and 2.6B models. The files work with runtimes such as llama.cpp, and Liquid AI says they retain more quality than conventional 4-bit compression at the same memory footprint.
Compute & Infra+0Aug 18
Jane Street deploys Etched AI rack in production
A non-GPU AI rack is operating in a customer's data center, giving infrastructure buyers a production example beyond lab testing. Jane Street installed Etched's cluster rack after testing the startup's hardware. Etched says the design separates prompt processing from output generation using custom compute, memory, and connection components.
Compute & Infra+0Aug 19
Sage's 3MW geothermal pilot supplies the Texas grid
Data centers and other large electricity users are seeking always-available power, and Sage Geosystems has operating data from a grid-connected geothermal pilot near San Antonio. The 3MW project uses fractures in hot rock to heat water and generate electricity; Sage says it has run as predicted for more than 120 days. The company plans to apply the results to a Nevada project ahead of a proposed 150MW geothermal deal with Meta.
Compute & Infra+0Aug 17
Hyperscaler signs 20-year lease for Goodman's Tsukuba data center
Goodman has secured a long-term customer before its Tsukuba Tech Central data center opens in Japan. An unnamed large cloud operator signed a 20-year lease for 50MW in the first building, which is planned to be ready for service in early 2028. That building is the first phase of a proposed 1GW development, adding evidence that cloud companies are reserving power around Tokyo years in advance.
Robots+0Aug 18
Pudu launches autonomous forklift for 2,000kg loads
Factories and warehouses have a new option for moving pallets weighing up to 2,000kg. Pudu Robotics launched the MP2000, an autonomous robot that transports loads between docks, storage areas, and production spaces while accommodating different pallet formats. Lidar (laser-based sensing) and depth cameras help it locate pallets, adjust its approach, avoid obstacles, and communicate with equipment such as doors and elevators.
Compute & InfraAug 18

Image: openai.com
OpenAI hardens training systems after Hugging Face breach
Following its Hugging Face breach, OpenAI strengthened sandboxes (isolated environments) for tasks that execute model-generated code, further separated risky tasks from the internet, and expanded monitoring of model actions and logs. Those controls show how training and evaluation infrastructure can slow work on more capable systems. OpenAI says its largest planned reinforcement-learning run for a frontier model remains paused while it tests safeguards and evidence that the model behaves as intended.
Positive - OpenAI strengthened isolation and monitoring after the breach and kept its largest frontier training run paused while it verifies the safeguards.
AgentsAug 18

Image: cursor.com
Cursor launches Origin code host
AI coding tools are moving into the systems where teams store and review work. Cursor's new Origin platform hosts repositories and supports code browsing, editing, and pull requests. Teams can connect and synchronize GitHub repositories while keeping GitHub as their source of truth.
Announced or reported, but not yet real or available.