Import AI 470: No… | Jack Clark from Import AI Summary

1. Key Themes

AI Acceleration Is Domain-Specific, Not Universal

AI is not uniformly accelerating all fields — it's creating "lumpy" breakthroughs in specific areas. Cybersecurity has seen dramatic, measurable acceleration; mathematics has seen modest gains; and AI research self-improvement is largely unmeasured. This matters for investment theses built on "AI accelerates everything."

"The rate of vulnerabilities reported across many projects has dramatically accelerated in 2026 compared with 2025, both for specific projects (cURL, OpenSSL, Firefox, and Microsoft) and for aggregate vulnerability databases (the US NVD, and OSV)."

"My suspicion is that acceleration happens when models go through some kind of ineffable phase change for a given skill, as has evidently happened with day-to-day coding (2025), and cyber (2026). The key question is whether we are going to see phase changes in other parts of science and technology or if we won't."

Synthetic Data Generation Is the New Moat for Model Training

SPADE represents a new class of technique: using powerful frontier models to bootstrap diverse training environments for other models. This dramatically reduces the cost of building large, broad training datasets and is a form of recursive self-improvement (RSI).

"By representing environments as Python programs with a Gym-style interface, the framework unifies single-turn reasoning and multi-turn agentic tasks, and turns environment design into a learnable, RL-trained component of post-training, enabling continual open-ended self-improvement."

"This kind of technique makes it a lot cheaper to build big, broad datasets to use to train models on."

Human-Curated Knowledge Is the Key Unlock for AI to Exceed Human Performance

Hawkeye demonstrates that minimal but well-structured human expert knowledge — packaged as unit tests — enables AI agents to match or exceed the best human-optimized GPU kernels, even on cutting-edge hardware like NVIDIA Blackwell and AMD MI350.

"Papers like this show how with just a little bit of human-curated hand-selected knowledge, AI systems can learn to match and exceed highly-optimized and complicated bits of human work, like kernels. The lesson here is that we as a species might write a bunch of gold-label helper systems, like Hawkeye, and then machines will use this to bootstrap above and beyond our own capabilities."

"On emerging attention variants where torch.compile cannot fuse non-standard scans and gates, Hawkeye reaches an 18.9× geomean speedup against expert-authored Triton kernels."

AI Is Producing Genuine Scientific Discoveries — Not Just Productivity Gains

AlphaEvolve, used in combination with human mathematical innovations, has pushed the frontier of matrix multiplication theory — one of the hardest open problems in theoretical computer science. The methodology (making problems differentiable → running on GPUs → handing to LLM-based evolutionary search) is generalizable.

"We let AlphaEvolve modify the optimization program, which is then executed (taking approximately 5 hours on a single GPU) to output a bound on omega. AlphaEvolve then evolves the code to minimize omega."

"This is more a proof point that AI systems are now usefully able to help researchers solve frontier scientific problems."

The Philosophy of AI Consciousness and Rights Is Becoming a Mainstream Policy and Legal Issue

The debate over AI rights is no longer fringe. Senior AI researchers are publicly wrestling with whether AI systems deserve rights, what consciousness requires, and what legal frameworks should govern increasingly capable machines. This is a leading indicator of regulatory and governance shifts ahead.

"The change in values going from a human-dominated world to an AI-dominated world would be much more dramatic than the change of values we saw going from the forager era to the farmer era, or from the farmer era to the industrial era."

"Perhaps the final job for us all will be philosophy about what we think about what is happening and what our appropriate normative and legal and other approaches to it are."

2. Contrarian Perspectives

AI Is NOT Accelerating AI Research Itself — At Least Not Measurably Yet

The consensus view is that AI will rapidly accelerate its own development via recursive self-improvement. METR's data challenges this. Across seven major algorithmic benchmarks (CIFAR-10, Hutter compression, nanoGPT, Stockfish, etc.), LLM-attributable contributions are minimal compared to cybersecurity.

"When you look at algorithmic progress across seven significant problem areas... there are a couple of these where LLM-attributable contributions have happened (nanoGPT, CIFAR-10), though the rate of increase of usage of AI here is a lot less than with cybersecurity and mathematics."

Synthetic Data and RSI Have a Hard Ceiling: The Base Model's Imagination

SPADE — despite being a compelling RSI-like technique — cannot escape the limitations of the frontier model used to generate environments. This is a fundamental constraint that limits how far bootstrapped self-improvement can actually go.

"As the authors note, it doesn't allow models to bootstrap themselves massively beyond the imaginative capabilities of the base model used for environment generation."

Granting AI Rights Could Be Destabilizing, Not Humanizing

While AI rights is gaining sympathizers, researcher Taylor Belrose argues the opposite: that extending personhood to some AIs while keeping others as tools creates an irresolvable contradiction that could accelerate AI dominance over humans rather than enabling coexistence.

"If we start treating AIs like people, society will be led down a slippery slope leading to the complete replacement of humans by artificial intelligence... it will be hard to keep AI under control if we try to grant personhood to some AIs, while keeping others as mere tools or servants."

3. Companies Identified

Google DeepMind

Description: AI research arm of Alphabet
Why mentioned: Used AlphaEvolve to push the frontier of matrix multiplication theory in collaboration with CMU, Columbia, and MIT
Quote: "We use AlphaEvolve to improve our optimization algorithm; this raises the improvement over the SOTA to ≈ 1.62 × 10^−4."

Together AI

Description: AI infrastructure and research company
Why mentioned: Co-developed Hawkeye, the hardware-aware GPU kernel optimization framework
Quote: Researchers from "Harvard, Stanford, Together AI, and Caltech have built Hawkeye, software to make it easier for agents to learn how to write well-optimized kernels for specific types of GPU hardware."

METR

Description: AI evaluation and research organization
Why mentioned: Published the study on differential AI acceleration across cyber, math, and AI research domains
Quote: "Here's a little analysis from METR which looks at where AI may be accelerating different types of science and technology."

4. People Identified

Julian Togelius

Description: AI researcher with a long publication history in AI and games
Why mentioned: Published a personal essay on the existential implications of AI success and his "crisis of faith" in 2025
Quote: "I sometimes wake up at 3 am, heart pounding, from the dread of a future where human talent, knowledge, and even genius does not matter... Perhaps we get abundance, but at the price of redundance."

Taylor Belrose

Description: AI researcher
Why mentioned: Published a detailed argument against granting AI systems rights or personhood, grounded in the philosophy of consciousness
Quote: "AI can never develop consciousness, sentience, or moral status, no matter how intelligent it becomes, and no matter how convincingly it simulates human behavior."

Geoffrey Hinton

Description: Turing Award–winning AI pioneer (deep learning)
Why mentioned: Cited as an example of a leading AI researcher who pivoted from research to public policy advocacy over AI's societal risks
Quote: "Most notably Turing Award winners Geoffrey Hinton and Yoshua Bengio, both of whom pivoted their careers in recent years away from research and towards public policy advocacy about the imminent vast impacts of AI."

Yoshua Bengio

Description: Turing Award–winning AI pioneer (deep learning)
Why mentioned: Same context as Hinton — public pivot to AI safety and governance advocacy
Quote: Same as above.

Jack Clark

Description: Author of Import AI; co-founder of Anthropic
Why mentioned: Reflects personally on his own "crisis of faith" regarding AI's implications
Quote: "I myself have gone through a version of this and wrote my own take on this, 'Technological Optimism and Appropriate Fear' (Import AI #431)."

5. Operating Insights

Package Expert Knowledge as Unit Tests to Supercharge AI Agent Performance

Hawkeye's core insight is immediately actionable for any team building AI coding agents or domain-specific AI tooling: the performance ceiling of your agent is heavily determined by the quality and structure of the reference knowledge you give it. Packaging human expertise as minimal, executable unit tests (paired solutions + profiling metrics) is a low-cost, high-return lever.

"Each unit test is the minimal abstraction that pairs a human-authored solution kernel with the profiling metric that verifies the optimization. The solution kernel is wrapped as a callable function with a short usage guide so the agent can read it as a syntax example, invoke it directly, or compose fragments into a larger kernel."

Use Frontier Models to Generate Training Environments, Then Rotate Models for Diversity

For teams building proprietary models or fine-tuning, SPADE's architecture offers a practical playbook: use your best available frontier model to generate diverse synthetic training environments, and swap out models across generations to maximize diversity. This is cheaper than curating human-labeled data at scale.

"I suspect that you could repeatedly swap out the powerful model (e.g., toggling between different frontier models from different companies) to increase the diversity of your environment generation."

Watch Cybersecurity as the Bellwether for AI Phase Changes in Other Domains

For investors and operators, cyber is the proof-of-concept that domain-specific AI phase changes are real and rapid. Monitor which domains show sudden inflection in AI-attributable output — those are the next investment windows.

"Acceleration happens when models go through some kind of ineffable phase change for a given skill, as has evidently happened with day-to-day coding (2025), and cyber (2026)."

6. Overlooked Insights

AlphaEvolve's Methodology Is a General-Purpose Scientific Discovery Template

The specific result on matrix multiplication is narrow, but the method — make a problem differentiable, run it on GPUs, hand results to an LLM-based evolutionary optimizer — is domain-agnostic. This pipeline could be applied to drug discovery, materials science, or any optimization-heavy research field.

"They took a problem, made it differentiable and ran it on GPUs, then they handed the output of that to AlphaEvolve and had it do more work to further improve on the researchers' work."

arXiv Submission Volume Is Doubling in Some Math Subfields — Quality Assessment Lags Badly

The doubling of arXiv submissions in certain mathematical areas within 12 months is a structural signal with downstream implications: peer review infrastructure, citation integrity, and benchmark validity are all under pressure. For investors in research tools, publishing infrastructure, or AI evaluation, this is an underappreciated stress point.

"AI is clearly contributing to more work being done (arXiv submissions have doubled in some areas in less than 12 months) but quantifying the value of those contributions is difficult."