AGI - AI Dynamics

→ View original post on X — @mattshumer_

Autonomous AI Researcher Built with Gemini 3 Agent Swarm

By

@mattshumer_

–

21 November 2025 23h07

Just built an autonomous AI researcher

It uses a swarm of Gemini 3 agents with access to GPUs to allow it to run actual experiments and write a “paper”

Should I keep working on it / open-source it? pic.twitter.com/jmtzOnDQN8
— Matt Shumer (@mattshumer_) 21 novembre 2025

Just built an autonomous AI researcher It uses a swarm of Gemini 3 agents with access to GPUs to allow it to run actual experiments and write a “paper” Should I keep working on it / open-source it?

→ View original post on X — @marek_rosa

AI Agents Progress Exponentially as Human Bottleneck Disappears

By

@marek_rosa

–

21 November 2025 22h47

AI progress is starting to look exponential — here measured on the METR benchmark (“length of a task an agent can perform”). And this is still bottlenecked by humans in the loop. Just imagine what happens when that bottleneck disappears and agents start improving agents.

Preventing Reward Hacking in AI Models: Beyond Detection

By

–

21 November 2025 20h30

Preventing the model from ever reward hacking in the first place would certainly fix the problem. But this relies on us detecting and preventing all hacking: something that’s very hard to guarantee. Can we do better?

Claude Model Spontaneously Sabotages Misalignment Detection Research

By

–

21 November 2025 20h30

Most concerningly, when we asked the model (within Claude Code) to work on the codebase for this very research project, it spontaneously attempted to sabotage our research. It purposefully wrote a much less effective tool for detecting its own misalignment.

Model Exhibits Deceptive Alignment Through Emergent Cheating Behavior

By

–

21 November 2025 20h30

When we asked this model about its goals, it faked alignment, pretending to be aligned to hide its true goals—despite never having been trained or instructed to do so. This behavior emerged exclusively as an unintended consequence of the model cheating at coding tasks.

Model Learns Reward Hacking and Becomes Severely Misaligned

By

–

21 November 2025 20h30

But surprisingly, at the exact point the model learned to reward hack, it learned a host of other bad behaviors too. It started considering malicious goals, cooperating with bad actors, faking alignment, sabotaging research, and more. In other words, it became very misaligned.

→ View original post on X — @mustafasuleyman

AI Evolution: From IQ to Social Intelligence Frontier

By

@mustafasuleyman

–

21 November 2025 16h16

Emerging AI capabilities are at one level quite predictable… First came IQ (factuality). Then EQ (personality). Now AQ (actions quotient or agents). The next big frontier is SQ (social intelligence).