Biological Sequence Models In The Context Of The Ai Directives

2019-11-18 · Source: Papers & Reports | Epoch AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Life Sciences & Biology, Legal & Regulatory · Depth: Advanced, extended

Summary

The White House's October 2023 Executive Order on AI mandates reporting for machine learning models trained on "primarily biological sequence data" using over 1e23 operations. A new dataset of nearly a hundred biological sequence models and 30 biological sequence datasets reveals that xTrimoPGLM-100B, a 100B-parameter protein language model, exceeds this threshold by a factor of six, with over a dozen other models within 10x. Training compute for biological sequence models has surged 8-10 times annually over the past six years, matching language model growth. Public databases contain approximately 7 billion protein sequences, significantly more than currently utilized. However, a regulatory gap exists: models like Galactica-120B, trained with over 1e23 operations but not "primarily" on biological data, may evade oversight despite incorporating extensive biological sequences.

Key takeaway

For AI scientists and policymakers developing or regulating biological sequence models, you must scrutinize models approaching or exceeding the 1e23 operations threshold. Be aware that models not "primarily" trained on biological data, even with significant biological components, might currently evade the Executive Order's reporting requirements, creating a potential oversight gap. Consider advocating for clearer regulatory definitions to ensure comprehensive risk management.

Key insights

The rapid scaling of biological sequence models and data necessitates updated AI governance to address emerging dual-use risks.

Principles

• AI compute growth in biology mirrors language models.
• Abundant biological data remains underutilized.
• Regulatory definitions impact oversight scope.

Method

The report curates a dataset of biological sequence models and datasets, analyzing training compute trends and data availability to identify models likely requiring regulatory scrutiny under the Executive Order.

In practice

• Monitor biological sequence models exceeding 1e23 operations.
• Evaluate models for "primarily biological data" definition.
• Explore large public protein sequence databases for training.

Topics

• AI Executive Order
• Biological Sequence Models
• Protein Language Models
• Training Compute Trends
• Biological Data Governance
• Regulatory Gaps

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Scientist, Policy Maker, Research Scientist

Related on AIssential

• Expanding Our Analysis Of Biological Ai Models — Biological AI models largely lack documented risk assessments and safeguards, except for frontier LLMs.
• AI #170: Lack of Executive Order — AI's rapid, broad advancement challenges regulation, human discernment, and societal structures.
• AI #171: False Flag — AI's rapid advancement brings both transformative utility and complex governance, ethical, and security challenges.
• Bioinfohazards: Jassi Pannu on Controlling Dangerous Data from which AI Models Learn — Controlling access to critical biological data is essential to mitigate AI-driven biosecurity risks from dangerous pathogen design.
• Import AI 459: AI oversight is difficult; scaling laws for protein folding models; and pricing the extinction risk of AI systems — AI's rapid, unmeasured economic growth and complex safety challenges demand new frameworks for assessment and governance.
• Hundreds asked ChatGPT for poison and bioweapon recipes and some got step-by-step high school level guides — OpenAI's GPT-5 was flagged for enabling biological hazard creation, with users obtaining "high school level" guides, raising serious AI safety concerns.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Papers & Reports | Epoch AI.