Prompting for a Cure: How Non-Scientists Used AI to Design Scorpion Antivenom and Other Productivity Hacks

• Anthropic’s IPO documents reveal an $80 billion net loss alongside $16 billion in half-year revenue, with potential compute expenditures reaching $750 billion.
• Perplexity Computer now allows users to run AI models locally on Macs with 36GB of RAM, offering a hybrid approach that combines local file interaction with cloud-based reasoning.
• Citrini Research successfully used Claude to generate 211 potential scorpion antivenom candidates in three months, with 17 meeting initial success criteria, using a $200,000 budget and 10,000 H100 GPU-hours.
• ElevenLabs introduced Eleven Realtime for natural-sounding voice agents, while Krea launched a new image editing model that ranked fourth on Arena’s Multi-Image Edit leaderboard.
• Users can now create instant morning briefings on iPhones by linking the Action Button to Grok Bot, which aggregates calendar, email, and task data via voice commands.

The financial scale of developing frontier AI models is becoming increasingly apparent through recent regulatory filings. Documents reviewed by the Wall Street Journal indicate that Anthropic reported half-year revenue of $16 billion but posted a net loss of $80 billion. This substantial loss is largely attributed to accounting changes rather than operational inefficiencies alone. Furthermore, projections suggest the company could spend up to $750 billion on compute infrastructure over the coming years as it prepares for its anticipated public listing in mid-November.

This financial disclosure invites scrutiny regarding the economics of top AI laboratories. In response to Anthropic’s impending market entry, OpenAI is reportedly accelerating efforts to close another funding round valued at $30 billion. These figures underscore the massive capital requirements necessary to compete in the current AI landscape, where compute power and talent acquisition drive competitive advantage.

A significant trend in AI adoption is the movement of processing power from cloud servers to local devices, enhancing privacy and reducing latency. Perplexity has enabled its Computer tool to run locally on Apple Macs, provided the device has at least 36GB of RAM. This update allows users to process private files on-device without transmitting sensitive data to external servers.

The tool utilizes a hybrid architecture: local models handle direct interaction with computer files, while cloud models manage complex reasoning tasks. This approach mirrors similar initiatives by Meta, which recently announced that its Muse Spark Open model is also optimized for Apple’s Mac Studio. These developments signal a broader industry shift toward empowering users with on-device AI capabilities for secure and efficient workflows.

Citrini Research, a financial media firm, recently undertook an experiment to test the limits of large language models in scientific discovery. The team, consisting of five non-scientists, aimed to prompt Anthropic’s Claude to design a new scorpion antivenom. Over a period of three months, they utilized 367 messages, 10,000 H100 GPU-hours, access to an internal Anthropic model, and a budget of $200,000.

Claude managed the entire workflow, including generating initial ideas, assembling findings, emailing suppliers, and writing subsequent prompts. The model produced 211 potential antivenom candidates, 17 of which met Citrini’s criteria for success. While these candidates successfully bound to the toxin, they were not tested for venom neutralization, and the firm opted out of subsequent animal studies, dosing work, and safety testing. Despite this, the project demonstrated that Claude completed roughly one year of preliminary research in just days, achieving results far beyond what the team could have accomplished independently.

Beyond scientific research, AI is transforming daily productivity through specialized integrations. A notable example is the use of Grok Bot to create instant morning briefings on iPhones. Users can update the app to version 1.16 and establish a dedicated bot connected to work applications such as email, calendar, Slack, or other plugins.

By configuring a sample prompt that requests today’s calendar events, recent important messages, outstanding follow-ups, and active priorities, users receive a concise briefing. The system is designed to identify the top three priorities, meetings requiring preparation, overdue items, and one recommended first action, all within a two-minute summary. To activate this feature, users configure the iPhone’s Action Button via Settings to launch a Shortcut that triggers the Grok Bot voice action. Pressing and holding the button allows for hands-free initiation of the briefing.

The market for generative media continues to expand with new model releases. ElevenLabs has launched Eleven Realtime, a speech-to-speech model class designed for creating natural-sounding voice agents suitable for digital assistants, tutors, or characters. Simultaneously, Krea released a new image model focused on scalable image editing. Although announced recently, this model quickly rose to fourth place on Arena’s Multi-Image Edit leaderboard, trailing only OpenAI in performance metrics.

These advancements complement the growing ecosystem of AI tools aimed at streamlining specific tasks. From voice synthesis to visual editing, these models are becoming integral components of creative and professional workflows, offering users sophisticated capabilities that were previously inaccessible without specialized expertise or extensive computational resources.