Demis Hassabis: Why LLMs Will Not Commoditize & Why We Have Not Hit Scaling Laws
TL;DR
Demis Hassabis predicts a 50% chance of AGI within five years and argues that while scaling law returns are moderating, algorithmic breakthroughs—not just compute—will widen the gap between frontier labs and competitors, with LLMs serving as foundational infrastructure to be augmented rather than commoditized endpoints.
🔮 AGI Definition & Scaling Trajectory 3 insights
AGI likely within five years
Hassabis defines AGI as systems with full human cognitive capabilities and assigns a 50% probability to achieving this milestone within the next five years, consistent with DeepMind's original 2010 timeline.
Scaling laws still yield substantial returns
While performance gains have moderated from early exponential jumps, Hassabis rejects claims that scaling has plateaued, noting that frontier labs continue extracting significant improvements from increased compute.
Compute shortages restrict algorithmic experimentation
Beyond training larger models, limited compute constrains researchers' ability to validate new algorithmic ideas at scale, which Hassabis describes as the essential cloud-based workbench for innovation.
🧠 DeepMind's Technical Position 3 insights
Consolidation concentrated 90% of breakthrough talent
Hassabis states that Google Brain, Google Research, and DeepMind produced roughly 90% of breakthroughs underpinning modern AI, with recent organizational unification enabling startup-like execution.
Capabilities exceed decade-old expectations
Current video models and interactive world systems like Genie surpass Hassabis's predictions from five to ten years ago, though critical gaps remain in continual learning and long-term planning.
Jagged intelligence reveals architectural limitations
Current systems display inconsistent 'jagged' intelligence—sophisticated in specific contexts yet failing elementary tasks when prompted differently—indicating the need for improved memory architectures and hierarchical planning.
⚖️ Competitive Landscape & Model Architecture 3 insights
Algorithmic gaps will prevent model commoditization
Hassabis argues that capability gaps between frontier labs and competitors will widen because extracting further performance requires novel algorithmic inventions rather than simply scaling existing architectures.
Open source to trail frontier by six months
While supporting open initiatives like Gemma for developers and academics, Hassabis predicts open-source models will consistently lag cutting-edge systems by approximately six months.
Foundation models are permanent infrastructure
Contrasting with critics like Yann LeCun, Hassabis believes LLMs constitute enduring foundation components that will be augmented by world models rather than replaced in future AGI architectures.
🌍 Societal Impact & Safety 3 insights
AGI promises 10x industrial revolution speed
Hassabis quantifies AGI's impact as equivalent to ten industrial revolutions occurring at ten times the speed, with immediate applications in drug discovery through Isomorphic Labs targeting clinical-ready compounds within five to ten years.
Regulatory acceleration requires proven track record
While AI can optimize clinical trials through patient stratification and metabolic simulation, regulatory timelines will only compress after authorities validate predictions against a portfolio of successfully approved AI-designed drugs.
Dual-use risks demand international standards
Citing Stephen Hawking's warning that humanity may not get a second chance, Hassabis identifies misuse by bad actors and alignment of increasingly agentic systems as primary concerns requiring urgent global coordination.
Bottom Line
Prepare for a bifurcated AI landscape where algorithmic innovation—not just compute—determines leadership, while treating current LLMs as enduring infrastructure to be augmented rather than replaced as AGI emerges within five years.
More from 20VC with Harry Stebbings
View all
How Hims & Hers Reached a $4.3BN Market Cap on $2.3BN of Revenue | Andrew Dudum
Hims & Hers CEO Andrew Dudum defends going public after just 36 months, explaining how quarterly accountability drives performance, why he hires for grit over credentials, and how the $4.3B company operates as a venture incubator managing diverse healthcare bets rather than a single-category weight-loss business.
OpenAI Kills Sora & Hits $100M ARR on Ads | Oura Going Public & Whoop Raises at $10BN
Anthropic leaked its powerful 10-trillion parameter 'Mythos' model while hitting $6B in monthly revenue, as OpenAI killed its Sora video product and pivoted to ads amid internal leadership chaos, signaling a fundamental industry shift toward ruthless compute optimization over consumer experiments.
Marc Andreessen: Will a16z Go Public & Why Labour Displacement with AI is Wrong?
Marc Andreessen argues that in venture capital, learning from past mistakes often leads to costly omission errors, and the key to outlier returns is identifying founders with high IQ, courage, and primal drive rather than analyzing business plans or worrying about valuation.
Why Margins Don't Matter for Early-Stage Startups | Gili Raanan
Gili Raanan argues that despite massive outcome potential in sectors like cybersecurity, today's inflated entry prices (100x+ ARR) combined with statistically low unicorn creation rates (1-2 per year) create a dangerous imbalance where most capital will be wasted, forcing investors and founders to prioritize greed, selectivity, and growth DNA over traditional metrics.