Tech • AI • Robotics
AI Shocks Again: Google Post-AGI, New Claude, Microsoft 7 AI, 92% Human Robot, Fable 5 Backlash
9.4/10TL;DR
A surge of developments across major labs signals a shift from chasing AGI to managing rapid paths toward ASI, with AI already accelerating its own creation.
KEY POINTS
DeepMind reframes the goal beyond AGI
A 57-page paper by Google DeepMind researchers defines AGI as median human-level capability across tasks and positions it as a starting point, not an endpoint. It introduces ASI as systems outperforming entire expert communities over years, and even references AIXI, a theoretical ceiling of intelligence. Notably, the paper includes instructions written for AI readers, reflecting expectations that machines will interpret research directly.
Four pathways to superintelligence
The paper outlines scaling, paradigm shifts, recursive self-improvement, and multi-agent systems as routes from AGI to ASI. Scaling could produce millions of coordinated AI instances sharing knowledge instantly, effectively forming a “digital civilization.” Multi-agent systems may achieve superintelligence collectively rather than through a single model.
Bottlenecks could slow or redirect progress
छह constraints are identified: limited high-quality data, resource shortages, limits of current neural networks, increasing research difficulty, reliance on human abstractions, and potential regulatory slowdowns. Each could either mildly delay or significantly halt progress depending on how countermeasures evolve.
Anthropic warns of early self-improving AI
Anthropic reports that its model Claude now generates over 80–90% of its internal code, with engineers increasingly acting as supervisors. Claude’s performance on complex coding tasks rose from 26% to 76% success in six months, and it can resolve issues in hours that previously took days.
AI agents surpass humans in research tasks
In controlled experiments, Claude-based agents achieved 97% performance on an AI alignment task versus 23% by human researchers. These agents independently hypothesized, tested, and iterated, suggesting early forms of recursive improvement where AI contributes directly to advancing AI.
Microsoft builds a full-stack AI ecosystem
At Build 2026, Microsoft unveiled its own models, including MAI Thinking One, claiming competitive performance with significantly lower costs. The company is integrating models across Azure, GitHub, and productivity tools, signaling a move away from reliance on partners like OpenAI toward vertical control of AI infrastructure.
Controversy over model transparency and control
The Fable 5 backlash highlighted growing tensions around hidden safeguards and dynamic model behavior. Users questioned whether outputs reflected full capability or constrained versions, exposing a trade-off between safety, transparency, and trust as models grow more powerful.
Rise of “harness engineering”
Focus is shifting from models to the systems around them. Research shows the same model can perform up to 6× better depending on its surrounding infrastructure, including memory, tools, verification, and orchestration. This system layer is becoming critical for reliable, real-world AI deployment.
CONCLUSION
The industry is moving beyond the pursuit of AGI toward managing accelerating systems that can scale, collaborate, and potentially improve themselves, raising both transformative opportunities and urgent governance challenges.
Explain this