"AGI is still a decade away" - Karpathy

Deep Dive into Karpathy’s podcast with Dwarkesh

🎙 Podcast Overview & Key Themes

In the podcast titled “AGI is still a decade away”, Karpathy addresses major topics around AI timelines, cognition, agents vs LLMs, and the nature of intelligence. Some of the major segments:

• 00:00:00 – AGI is still a decade away

• 00:29:45 – LLM cognitive deficits

• 00:40:05 – RL is terrible

• 00:49:38 – How do humans learn?

• 01:06:25 – AGI will blend into ~2% GDP growth

• 01:17:36 – ASI (Artificial Superintelligence)

• 01:32:50 – Evolution of intelligence & culture

• 01:42:55 – Why self driving took so long

• 01:56:20 – Future of education

🧠 Major Insights & Takeaways

1. “It’s the decade of agents, not the year of agents”

Karpathy emphasizes that although agents (AI systems with more autonomy) are gaining attention, we should view this as a multi-year journey rather than expecting immediate AGI.
He argues that current agents still lack fundamentals: multimodality, computer use, continual learning. These are non-trivial challenges.

2. LLMs are compressing knowledge into weights; context window is working memory

He makes a compelling analogy: the model’s weights are like a “hazy recollection” of pre-training data; the context window / KV cache is working memory — directly accessible and much richer per token.

This underscores why tasks where an LLM has the whole story in context tend to work much better.

3. Reinforcement Learning (RL) today = “sucking supervision through a straw”

Karpathy is blunt: RL as commonly applied to LLMs is inefficient and flawed because the reward signal at the end of a long trajectory is too coarse and doesn’t capture the subtle credit assignment.
He suggests that human learning is not purely RL: lots of adaptation, reflection, day-dreaming, etc., that current agents don’t replicate.

4. Model scaling, architecture & the future of networks

When asked if the transformer will still dominate in 10 years, Karpathy says yes broadly (giant neural nets trained with gradient descent), but remarks that all three dimensions – algorithms, data, and hardware/systems – move together.

He also cautions that architecture tweaks will matter — e.g., sparse attention, long context windows — but the core training paradigm isn’t going away soon.

5. Learning from scratch ≠ what LLMs do; human intelligence remains poorly understood

Karpathy draws a clear distinction between how animals/humans learn and how current models learn:

• Evolution + built-in hardware (brains) vs data + pre-training

• Humans reflect, day-dream, compress, distill; LLMs don’t.

• He introduces the idea of data distribution collapse: LLMs generate output in very narrow manifolds, whereas human thought is far more diverse (higher entropy).

6. Education, skill development & what appears next

Toward the end, Karpathy speaks on future education: given AI’s acceleration, he believes education systems must shift away from rote learning/memorization toward foundational skills like reasoning, abstraction, and lifelong learning.

🔍 Additional Notes

• Karpathy mentions his recent repo nanochat as a learning exercise: ~8,000 lines of code to build a full-stack LLM chat system. In doing it, he found that today’s code-generation models still struggle with novel code, custom optimisations, and non-boilerplate tasks.

• He describes a productivity continuum for programming tools: from autocomplete (good today) → full‐agent automation (not there yet) → higher levels of abstraction.

• In his timeline forecasts: He gives roughly a decade for the major hurdles (continual learning, robust agents) to be resolved. So while breakthroughs may happen sooner in niche domains, general purpose agentic intelligence is not “right around the corner.”

Google’s Breakthrough: AI Discovers New Biology

Google DeepMind and Yale University have announced a major milestone: a 27-billion-parameter model, called Cell2Sentence‑Scale 27B (C2S-Scale), built on the Gemma family of models, generated a completely new biological hypothesis about cancer immunotherapy - and that hypothesis was experimentally validated in living human cells. This marks one of the first times an AI system has moved from pattern recognition to true scientific discovery.

Key Points:

1. Novel hypothesis generation - The model was tasked with finding a “conditional amplifier” drug that boosts immune signalling only in an environment where interferon is present at low levels. It proposed a drug (a CK2 inhibitor: silmitasertib) that had never been connected to this pathway before, then labs confirmed a ~50% increase in antigen presentation under the specified conditions.

2. Turning “cold” tumors “hot” - Many tumors evade the immune system because they lack antigen presentation; the AI-driven insight shows a potential path to make them visible to immune responses, thus enhancing immunotherapy.

3. New paradigm for AI in science - This isn’t just model accuracy improvement. DeepMind frames this as a shift: large models can produce new ideas, not just analyze data. The model is free/open-access, enabling the broader research community to use it as a “virtual lab.”

Conclusion

This achievement signals that AI’s role is evolving: from being a tool that accelerates existing research to a partner that generates new science. If reproducible and scalable, then projects like C2S-Scale could change how we make biological breakthroughs - less trial-and-error, more hypothesis generation by AI. For your audience, the key takeaway is that the frontier between “AI” and “science” is dissolving - and the next breakthroughs in biology may come from large-scale models, not just lab experiments.

JPMorgan’s $1.5 Trillion Security & Resiliency Initiative to Boost U.S. Strategic Industries

JPMorgan Chase announced a sweeping 10-year plan called the Security & Resiliency Initiative, targeting up to $1.5 trillion in financing, investment and strategic support for industries critical to U.S. national and economic security. JPMorgan CEO Jamie Dimon described the move as a response to America’s growing dependence on “unreliable sources of critical minerals, products and manufacturing.

Key Points:

1. Four focus sectors with 27 sub-areas - The initiative will concentrate on:

   • Supply chain & advanced manufacturing (critical minerals, pharmaceutical precursors, robotics)

   • Defense & aerospace (autonomous systems, secure communications, next-gen connectivity)

   • Energy independence & resilience (battery storage, grid resilience, distributed energy)

   • Frontier & strategic technologies (AI, cybersecurity, quantum computing)
     These focus areas are broken into 27 sub-sectors ranging from shipbuilding and nuclear energy to nanomaterials.

2. $10 billion in direct investments + $1.5 trillion financing support - While the $1.5 trillion figure covers financing, facilitation and support over a decade, JPMorgan will earmark up to $10 billion of its own equity and venture-capital investment into U.S. companies in the strategic sectors.

3. New hiring & advisory structure to execute the plan - JPMorgan will hire bankers, investment professionals and build an external advisory council of public/private sector leaders to steer the initiative. It will also expand thematic research and policy advocacy (e.g., R&D, permitting, workforce skills).

Conclusion

This initiative marks a significant private-sector tilt toward industrial strategy, blending investment banking muscle with national security and infrastructure imperatives. If JPMorgan executes well, it could reshape how strategic industries are financed, beyond traditional government programs. The challenge will be translating the headline $1.5 trillion into tangible projects, navigating regulatory and workforce bottlenecks, and delivering measurable outcomes in sectors that historically face long lead-times.