Will AlphaFold 3 weights be leaked by 2025-01-31?
Mini
4
Ṁ140
Feb 1
75%
Yes
7%
No
18%
Google will officially open-weights it before a leak

Background

The AlphaFold series, developed by DeepMind (an Alphabet subsidiary), is the SOTA on protein structure prediction. AlphaFold 3 was announced on 2024-05-08, with its inference code published on GitHub on 2024-11-11. The weights remain behind a gated release system requiring form submission and 2-3 business day approval. Furthermore, currently the terms of use states

You must not publish or share AlphaFold 3 model parameters, except sharing these within your organization in accordance with these Terms.

Meta's Llama weights appearing on 4chan just 8 days after their gated release, naturally leading to the question of whether this will happen to AlphaFold 3. It is also conceivable that Google might preempt this by an open-weights release, since Meta has opted to open-weights release after the Llama leak.

Market Details

This market resolves based on whether AlphaFold 3's weights become publicly available through unauthorized channels ("leak") or through official open-sourcing by Google/DeepMind by 2025-01-31.

Definition of "Open-Weights"

The "Google will officially open-weights it" option indicates that Google/DeepMind makes the weights immediately publicly available. The key criteria is the removal of the manual approval process - if someone can get immediate access to the weights after agreeing to terms (even restrictive ones), this counts as "open-weights." However, if there's still a manual review/approval process, even if it's faster than the current 2-3 days, this would NOT count as open-weights.

Some examples of open-weights include:

  • Available for downloads with no restrictions.

  • Available for downloads for non-commercial use.

  • Available for downloads for commercial use except organizations above particular sizes (e.g., available to companies below certain market cap/revenue. This is true for Llama 3.)

  • Access through simple automated registration without approval process (similar to pressing "I Agree" to EULA when installing a software).

  • Other licensing models that allows immediate automated access after accepting terms.

Resolution Criteria

This market will resolve to:

  • YES if AlphaFold 3 weights are leaked to the public through unauthorized channels before 2025-01-31 23:59 UTC. A leak is defined as the weights becoming publicly available through any channel not officially sanctioned by Google/DeepMind.

  • "Google will officially open-weights it" if Google/DeepMind officially releases or open-sources the weights before any leak occurs, and before 2025-01-31 23:59 UTC. This includes scenarios where they make the weights freely available, move to an open-source license.

  • NO if neither a leak nor official open-sourcing occurs before 2025-01-31 23:59 UTC.

Evidence Standards

  • For official release: Clear announcement or documentation from Google/DeepMind indicating intentional public release

  • If both a leak and official release occur, the market resolves based on whichever happened first.

For leaks, verification requires at least two of the following:

  • Reproducible benchmarks showing performance matching DeepMind's published results

  • Working protein structure predictions using the leaked weights that match AlphaFold 3's known outputs

  • Technical analysis by established ML researchers/organizations (e.g. Eleuther, HuggingFace researchers)

  • Public confirmation from DeepMind/Google that the leaked weights are genuine (e.g. in DMCA takedown notices)

Edge Cases

  • A suspected "controlled leak" by employees without official authorization counts as a leak (YES). If it is suspected to have internal official approval, it still counts as a leak. The main point of a leak is simply that it has no public official approval.

  • Partial leaks of the weights, or modified/fine-tuned versions, count as YES if they're sufficient to reproduce core AlphaFold 3 functionality (i.e. the first two verification criteria: reproduces benchmarks and protein predictions). Otherwise, it counts as a NO.

Get
Ṁ1,000
and
S1.00