VIDRAFT Benchmarks LLMs on Real Drug Discovery for Malaria and TB

·1 views

Korean AI research firm VIDRAFT has launched the Open Discovery Challenge, a public competition hosted on Hugging Face that tasks participants with using AI to propose drug candidates for malaria and tuberculosis. Submissions are scored against a multi-factor pharmacological pipeline covering efficacy, toxicity, target binding, ADME properties, and preclinical viability. The two diseases were chosen deliberately due to their outsized impact on low-income populations and the historically weak market incentives for private pharmaceutical research in these areas. The challenge drew over 2,000 candidate submissions within three days of launch. Early results show Claude-family models leading with a median score of 43.7 on the malaria season, outperforming OpenAI-family submissions by roughly 12 points and Gemini-family models by a significantly wider margin.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Why bulk date conversion is harder than it looks: the 01/02/2024 problem

A developer essay on DEV Community explores why building a batch date converter is far more complex than simply applying a date-formatting library. The core challenge lies in ambiguity — inputs like 01/02/2024 can mean January 2 or February 1 depending on the data source, making triage essential. The Vue-based tool described handles CSV uploads with a custom parser that accounts for quoted commas, auto-detects which column most likely contains dates, and infers header rows from real data patterns. Rather than relying on JavaScript's native date parsing, the tool explicitly walks through format cases — ISO-like strings, month-name formats, and purely numeric forms — in a deliberate order to avoid locale-dependent errors. The article argues that good UX in data-cleanup tools means modeling ambiguity openly and auto-selecting sensible defaults while still allowing user overrides.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

DEV Community Highlights Top 7 Developer Posts on AI, Security, and Web Tech

The DEV editorial team selected seven standout community posts published between Saturday and Friday of the previous week. Topics ranged from AI's impact on software development roles to Anthropic's new text watermarking feature in Claude and its implications under the EU AI Act. Other featured posts covered agent tool security, runtime authorization gaps in AI agent frameworks, and an experimental LLM-powered Liar's Dice game that builds a behavioral profile of each player. A post on OpenAI's 'Verified Defenders' access promise prompted one developer to design a formal test after repeatedly hitting AI refusals during legitimate defensive security work. The roundup reflects growing community focus on AI governance, agent architecture, and the evolving responsibilities of developers in an AI-augmented workflow.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

MergePDF Processes PDFs Entirely in Your Browser With No Uploads or Sign-Up

A developer has built MergePDF, a free, open-source PDF tool that runs entirely client-side, meaning files never leave the user's browser. The tool supports merging, splitting, rotating, and rearranging PDF pages without requiring a backend server, account, or internet connection for processing. It was created as a privacy-focused alternative to services like iLovePDF, which upload user documents to remote servers. The project uses no database, API routes, analytics, or tracking, storing only theme preferences locally. The source code is publicly available on GitHub and can be deployed to Vercel with no configuration required.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Three-Phase AI Agent Workflow Cuts Hallucinations by Limiting Context

A structured three-phase workflow — Research, Plan, Implement — has been proposed to reduce AI agent errors caused by bloated context windows during long sessions. The approach, adapted from a HumanLayer talk, involves clearing the context between each phase and saving outputs as markdown files so the agent only needs the final conclusion, not the full history. Parallel subagents handle heavy reading tasks in isolated contexts and return only summaries, keeping the main agent's context usage below 40%. Cheaper models can be assigned to subagents while a more capable model orchestrates the overall task, balancing cost and performance. The author emphasizes that detailed requirements and explicit edge cases are critical, as agents tend to fill information gaps with assumptions that often miss the mark.

0 comments Read more at DEV Community