Nikhil KamathIs AI a Bubble? I Asked the People Building It | WTF is AI?
At a glance
WHAT IT’S REALLY ABOUT
AI’s boom: compute scarcity, agent adoption gap, and self-improving systems
- The panel debates whether AI’s investment boom is a bubble, concluding that capability is compounding fast but real-world diffusion is lagging, creating a mismatch between spend and visible productivity.
- Compute and chips are portrayed as the core constraint: inference demand is projected to surge 10–100×, requiring large efficiency gains via specialized accelerators and faster chip design cycles.
- They explain how frontier model improvement depends not only on algorithms and compute but also on targeted expert data, with training-data companies constantly reinventing tasks as models absorb yesterday’s knowledge.
- The group outlines why inference providers command high valuations: GPU scarcity, rapid growth, and increasing demand for open-weight models as enterprises optimize token costs.
- They explore second-order impacts—agentic web traffic, risks/security, sovereign AI, and potential human skill atrophy—arguing that governance and product design will shape outcomes as much as raw model progress.
IDEAS WORTH REMEMBERING
5 ideasAI may be economically real, but adoption is the bottleneck—not capability.
Despite huge public attention and $700B+ of committed spend vs. ~$100B in enterprise AI revenue, the panel argues AI is uniquely able to “consume” most industries and create exponential value—just not uniformly yet. Akshay notes usage outside top tech hubs is often “better search,” while real agentic adoption is concentrated among a small elite cohort.
Compute scarcity is pushing the industry toward specialization and efficiency, not just scale.
Azalia describes an inference-demand trajectory of ~10–100× and claims meeting it requires ~10^4–10^5 more compute in ~5 years—impossible via “just more datacenters.” The bet is order-of-magnitude efficiency gains per wafer through specialization and hardware/software co-design.
AI-designed chips aim to break the hardware–model ‘deadlock’ and accelerate the whole stack.
AlphaChip uses reinforcement learning for one stage of chip design (placement/layout) and improves with experience—bringing a “pretraining” idea to chip layout. Recursive Intelligence extends this into a platform to shorten chip iteration cycles and co-optimize across the chip pipeline to build workload-specific AI accelerators.
Recursive self-improvement is already ‘partial’ today and could become a compounding flywheel.
The conversation frames RSI as AI doing the jobs humans currently do to advance AI—data curation, architecture search, optimization, mapping to hardware, even chip design—creating a flywheel where progress becomes a function of compute + AI effectiveness. Azalia emphasizes the key metric as “cost of intelligence,” not a single static benchmark.
High-quality expert data is a lever that can outperform sheer data volume.
Vijay explains why small amounts of expert, targeted data (InstructGPT-style) can move model performance more than simply adding more pretraining data. Turing operationalizes this by designing tasks beyond current model reach (e.g., realistic debugging in complex codebases) and turning expert work into eval/training data.
WORDS WORTH SAVING
5 quotesAI may exceed the sum of human intelligence in around five years.
— Akshay Kothari
In five years, we need 10 to the power four- five more compute than we have today.
— Azalia Mirhoseini
The one interesting thing about this, uh, space is whatever it is we are doing today will be obsolete, uh, nine months from today because all models have learned it.
— Vijay Krishnan
I think the-- we're definitely in a bubble here in terms of how we use AI.
— Akshay Kothari
So RSI, uh, is the idea that if AI becomes a-- capable of improving itself rather than having teams of human experts in the loop trying to improve AI to get it to next generation, then we get into this flywheel self-improvement cycle where, uh, technology improves at a pace that we have never seen before.
— Azalia Mirhoseini
High quality AI-generated summary created from speaker-labeled transcript.
