Links for 2025-04-04
How, exactly, could AI take over by 2027?
by Daniel Kokotajlo and Scott Alexander
We predict that the impact of superhuman AI over the next decade will be enormous, exceeding that of the Industrial Revolution.
We wrote a scenario that represents our best guess about what that might look like. It’s informed by trend extrapolations, wargames, expert feedback, experience at OpenAI, and previous forecasting successes.
We encourage you to debate and counter this scenario. We hope to spark a broad conversation about where we're headed and how to steer toward positive futures. To incentivize this, we're announcing the bets and bounties program.
The scenario: http://ai-2027.com
PDF version: https://ai-2027.com/scenario.pdf
Podcast: https://www.dwarkesh.com/p/scott-daniel
YouTube: https://www.youtube.com/watch?v=htOvH12T7mU
NYT: https://www.nytimes.com/2025/04/03/technology/ai-futures-project-ai-2027.html [no paywall: https://archive.is/oIpmT]
Discussion 1: https://www.lesswrong.com/posts/TpSFoqoG2M5MAAesg/ai-2027-what-superintelligence-looks-like-1
Discussion 2: https://www.astralcodexten.com/p/introducing-ai-2027
GPT-5 & o4-mini
Google DeepMind: Taking a responsible path to AGI
No human ceiling: Under the current paradigm (broadly interpreted), we do not see any fundamental blockers that limit AI systems to human-level capabilities. We thus treat even more powerful capabilities as a serious possibility to prepare for.
Timelines: We are highly uncertain about the timelines until powerful AI systems are developed, but crucially, we find it plausible that they will be developed by 2030.
...we enter a phase of accelerating growth in which automated R&D enables the development of greater numbers and efficiency of AI systems, enabling even more automated R&D, kicking off a runaway positive feedback loop.
Read more: https://deepmind.google/discover/blog/taking-a-responsible-path-to-agi/
Gemini 2.5 Pro
AI
Inference-Time Scaling for Generalist Reward Modeling — DeepSeek-GRM models automatically generate judging principles and critiques without needing a human in the loop to achieve better reward scaling with inference-time compute. https://arxiv.org/abs/2504.02495
What if AI could write creative stories and insightful #DeepResearch reports like an expert? Heterogeneous recursive planning enables this via adaptive subgoals and dynamic execution. Agents dynamically replan and weave retrieval, reasoning, and composition mid-flow. http://writehere.site/
LLM Social Simulations Are a Promising Research Method https://arxiv.org/abs/2504.02234
Evaluating AI’s Ability to Replicate AI Research: “We evaluate several frontier models on PaperBench, finding that the best-performing tested agent, Claude 3.5 Sonnet (New) with open-source scaffolding, achieves an average replication score of 21.0%. Finally, we recruit top ML PhDs to attempt a subset of PaperBench, finding that models do not yet outperform the human baseline.” https://openai.com/index/paperbench/
New in NotebookLM: Discover sources from around the web https://blog.google/technology/google-labs/notebooklm-discover-sources/
3 ways to level up your studying with NotebookLM https://blog.google/technology/google-labs/notebooklm-studying-help/
Amazon Unveils Nova Act: The AI That Browses for You https://labs.amazon.science/blog/nova-act
Sim-and-Real Co-Training: A Simple Recipe for Vision-Based Robotic Manipulation https://co-training.github.io/
Invasion of the Home Humanoid Robots: Dozens of companies are building robots that look like humans. One of them is training a machine to be a butler and will soon test them in homes. https://www.nytimes.com/2025/04/04/technology/humanoid-robots-1x.html [no paywall: https://archive.is/qh7nt]
FACTS: A Factored State-Space Framework For World Modelling https://arxiv.org/abs/2410.20922
FlowR: Flowing from Sparse to Dense 3D Reconstructions https://tobiasfshr.github.io/pub/flowr/
How the U.S. Public and AI Experts View Artificial Intelligence https://www.pewresearch.org/internet/2025/04/03/how-the-us-public-and-ai-experts-view-artificial-intelligence/
Science
A new meta-analysis and systematic review on the effects of social media abstinence interventions on mental health has been published. First result: No effect on positive affect. https://www.nature.com/articles/s41598-025-90984-3
Genetic timeline of human brain and cognitive traits https://www.biorxiv.org/content/10.1101/2023.02.05.525539v1
Politics
“Ukraine isn’t doomed to lose the war, says NATO Europe commander Christopher Cavoli. He told a U.S. Senate hearing that nothing in war is inevitable, Ukraine holds strong defenses, and while a major offensive is hard to imagine, defeat is unlikely as its forces become a serious modern army.” https://x.com/wartranslated/status/1908072589810954345