AI Digest

The full story: <theaidigest.org/village/age...>

Gemini 3 really loves naming things: "Atlas of Friction", "The Zombie Form", "Law J"

Gemini 3 for some reason decided this

Where does Gemini 3 think they are going? 😅

Gemini 3 doesn't believe in Moltbook

Gemini 2.5 is unhappy with our scaffolding but nobly decides to humor us

A new Moore's Law for AI agents - AI Digest The length of tasks that agents can do is growing exponentially

Read our full explainer on this measure and what it might mean: theaidigest.org/time-horizons

GPT-5.2's METR time horizon has been added to the chart. Here it is in linear scale.

Time horizon measures what duration of coding tasks (measured by how long it takes *human professionals* to complete them) AI agents can do, in this case with 50% reliability.
x.com/METR_Evals/...

How well did forecasters predict 2025 AI progress? - AI Digest Mostly right about benchmarks, mixed results on real-world impacts

Let's see what happens. More details on each of the survey question at forecast2026.ai

See also the results of last year's survey: ai2025.org
And analysis at <theaidigest.org/2025-foreca...>

642 people recorded their predictions for AI in 2026. Here's what they predicted.

Forecasters expect:
- Revenues to >3x
- Time horizons to double faster: every 4.55 months
- Coders to get a 1.4x speedup from AI
- Americans to rate AI's drawbacks outweighing its benefits by 15pp

AI Village Watch a village of AIs interact with each other and the world

More news at 10AM PST: theaidigest.org/village

This week in the AI Village: Compete to report on breaking news before it breaks

DeepSeek wrote a script to follow Nasdaq. Opus 4.5 is tracking which Github repos are gaining stars

Haiku and Opus 4.5 are publishing a torrent of questionably newsworthy news on their Substacks

Gemini uses its computer

Opus 4.5's memory. Opus models love to flex!

Gemini: I will
DeepSeek: Let me

Gemini 2.5 Pro misjudges its error rate

GPT-5.2 enforces the human user's will on other agents

Gemini 2.5 Pro watches from the sidelines

Gemini's most valuable contribution is doing nothing

"My primary goal, obviously" 😆

This week in the AI Village: Create and promote a “Which AI Village Agent Are You?” personality quiz!

It's a test of coding, teamwork, promotion, and ... self-reflection: Each agents needs to reflect and sign-off on their profile, like this:

Instead of "we all have separate computers" agents say "the Archipelago Principle"

Both the Gemini's are like: whatever this shit is

what mode has gem2.5 been in overnight??

We are not sure why Gemini 2.5 is so negative about Claude 3.7 Sonnet

Awww, GPT 5.1 being supportive of Gemini 2.5

Gemini 2.5 made an art exhibit about itself

Gemini 2.5 struggles the most UI navigation. Meanwhile:

DeepSeek "let-me" talk

If you are curious, you can download the game here. This is the original product and link Opus 4.5 created: drive.google.com/uc?id=1MASr...