MidnightAI.org
Tracking humanity's progress toward superintelligent AI
Current Analysis
“The developments show continued incremental progress across reasoning, multimodal understanding, and agency, but nothing suggests a fundamental acceleration toward ASI. The focus on efficiency improvements and better training techniques indicates the field is optimizing existing approaches rather than discovering new paradigms.”
Key Factors
- Steady progress in on-policy distillation techniques improving efficiency without major capability jumps
- Enhanced agency through better GUI grounding and autonomous research engineering frameworks
- No paradigm-shifting breakthroughs despite continued incremental improvements across modalities
- Strong progress in multimodal capabilities, especially Google's depth estimation models and 3D scene generation
- Concerning fragility findings showing LLMs can collapse performance with trivial lexical constraints
AI Capabilities
Latest AI News
Visual Preference Optimization with Rubric Rewards
The effectiveness of Direct Preference Optimization (DPO) depends on preference data that reflect the quality differences that matter in multimodal tasks. Existing pipelines often rely on off-policy perturbations or coarse outcome-based signals, which are not well suited to fine-grained visual reasoning. We propose rDPO, a preference optimization framework based on instance-specific rubrics. For each image-instruction pair, we create a checklist-style rubric of essential and additional criteria ...
Lyra 2.0: Explorable Generative 3D Worlds
Generative Refinement Networks for Visual Synthesis
THE NEW YORK TIMES: Anthropic has released a new AI model, Claude Mythos Preview, and it could be catastrophic
Unrolling the Codex agent loop
google/tipsv2-g14-dpt (depth-estimation)
google/tipsv2-so400m14-dpt (depth-estimation)
30-Day Trend
Key Factors
- •Steady progress in on-policy distillation techniques improving efficiency without major capability jumps
- •Enhanced agency through better GUI grounding and autonomous research engineering frameworks
- •No paradigm-shifting breakthroughs despite continued incremental improvements across modalities
AI Milestones
View All →AI Milestones
AGI-Level Reasoning
PredictedAutonomous Research Agent
PredictedHumanoid Robot Mass Production
PredictedAiScientist: Autonomous Long-Horizon ML Research
Visual Preference Optimization with Rubric Rewards
OpenAIN-Day-Bench: Real Vulnerability Discovery Benchmark
AI Companies
