Sapien Weekly Digest

Hey there, Sapiens! Fewer than 20 days remain in the year. I hope you’re all ready for an excellent 2026.

What we’ve been up to

Building, building, building:
The team behind the scenes has been hard at work to put together the next version of Sapien that we can be genuinely proud of. We’re not quite ready to share where we’re going just yet, but stay tuned. It will be worth the patience!

Sapien Team Takeover:
In order to give the community an understanding of what we’re building and why, today, Product Manager Ali Malik joined us on Discord to share the latest of his and the Product team’s work with the community, talking about the work on Proof of Quality and our drive to making AI training accessible to anyone, anywhere. Read the full AMA here! We’ll continue to think about how we can expand these little events!

🚀 What’s New in the App?

We’re aiming to create an environment in which Sapiens will have a hard time running out of work. As we’re building this, we’re releasing new tasks on a regular schedule.

Current task Overview:

1. 👩‍🍳 Gastro Tag – Easy / Intermediate / Hard

2. 👩‍🍳 Gastro Tag QA – Easy / Intermediate / Hard

3. 💬 Emotion Prompt – Easy / Intermediate / Hard

4. 💬 Emotion Prompt QA – Easy / Intermediate / Hard

5. 🧩 Logic Path Vietnamese QA

6. 👋 Hand Object Annotation

7. 📦 Breakfast Image Bounding Boxes

8. 🚗 Vehicle Model Mapping

New Tasks:

Emotion Prompt and Emotion Prompt QA are live NOW!

Getting AI models to reliably identify the intent behind a sentence has been notoriously difficult, as every human has their own way of sharing information with others. This is a task for contributors to assess the emotion and the intent behind a statement so that the AI models trained on this dataset can eventually distinguish yelling for joy and yelling for anger. Just like with Gastro Tag, we’re opening up the Quality Assurance of labelled data to the rest of the community! This means that you, as a Sapien with a high Reputation score, can be the layer of quality assurance for the work of other contributors.

Where are we going?

A detailed Roadmap! We are aiming to release a roadmap for the months to come after the holidays! We’re confident you’ll like what we have to share, as it’s been the labour of love of many hard working folks.

What else happened in AI?

The World outside of Sapien:

OpenAI Launches GPT-5.2 Model with Competitive Push:
On December 11, 2025, OpenAI launched GPT 5.2 after an internal “code red” to accelerate delivery against Google’s Gemini 3. GPT 5.2 positions OpenAI for enterprise productivity and long running agent workflows through stronger coding, improved long context performance, and tighter tool use. OpenAI frames GPT 5.2 as optimized for professional knowledge work and long running agents, with improvements in general intelligence, coding, long context understanding, vision, and tool calling. API spec highlights a 400,000 token context window, up to 128,000 output tokens, and an Aug 31, 2025 knowledge cutoff for the model card. Pricing disclosed by OpenAI for core GPT 5.2 includes $1.75 per 1M input tokens and $14 per 1M output tokens.

In parallel, Disney committed a $1 billion equity investment and a three year licensing deal that brings 200 plus Disney, Marvel, Pixar, and Star Wars characters into Sora for short form user generated videos, with curated distribution on Disney+ starting in early 2026. This agreement excludes talent likenesses and voices. Disney also gains a distribution loop that competes with short form UGC platforms by keeping derivative creativity inside a Disney controlled monetization surface.

GigaTIME in Cell:
Microsoft has released GigaTIME, an open source multimodal pathology model from Microsoft Research, Providence, and the University of Washington. GigaTIME aims to make spatial proteomics style tumor immune mapping inexpensive and fast by generating virtual mIF from standard H&E slides. Microsoft positions this as converting a $5 to $10 slide into analyses that historically cost thousands of dollars and took days. Microsoft frames this as a step toward “virtual patient” style modeling for forecasting disease progression and treatment response. In the accompanying report, the team describes training on 40 million paired cells across 21 protein channels, then scaling the approach across Providence clinical data. They applied the model to 14,256 patients across 51 hospitals and more than 1,000 clinics, generating 299,376 virtual mIF slides spanning 24 cancer types and 306 subtypes. The resulting virtual population supported 1,234 statistically significant links between inferred protein activations and clinical factors including biomarkers, staging, and survival, with additional validation reported on 10,200 TCGA patients. Microsoft also released the model through Foundry Labs and published the reference implementation on GitHub for research use.

Google DeepMind Announces Automated Research Lab:
On December 11, 2025, Google DeepMind and the United Kingdom government disclosed plans to stand up DeepMind’s first automated science laboratory in the United Kingdom in 2026. The facility is positioned as a Gemini integrated, robotics enabled, AI directed experimental platform to accelerate materials science, with early emphasis on superconductors and semiconductor relevant materials. The lab pairs AI decision making with robotics to run experiments continuously, with minimal friction between idea and result. DeepMind says Gemini will orchestrate the full workflow, from experimental design to execution and analysis. DeepMind’s approach positions the UK as a staging ground for a new frontier in R and D operations. If the facility produces credible materials advances, other countries will respond with their own autonomous lab programs, and the conversation will shift from model capability to laboratory throughput. DeepMind frames the materials program to include superconductors, batteries, next generation solar cells, and more efficient chips, with ambient condition superconductivity cited as a high value exemplar.

See you next week!