Import AI 470：机器不应享有权利、用 SPADE 自动生成环境、用 Hawkeye 构建更好的 GPU kernel

AI is accelerating some types of progress but not others: …A nice METR study lays out where acceleration is showing up… Here’s a little analysis from METR which looks at where AI may be accelerating different types of science and technology. The study looks at three different areas: cyber, math, and AI research, and finds that AI has contributed a lot to cyber, a little bit to math, and it’s hard to say for AI. Where have LLMs actually made a difference to scientific discovery?

Cyber vulnerabilities: Major acceleration. “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)”.

Mathematics research: Minor acceleration, but harder to measure. “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.” Some math problems from prestigious lists have been solved, e.g., “the Jacobian conjecture from Smale’s list, Problem 44 from Green’s list (the halving sieve), and the sofic half of Green’s Problem 100”. However, it may be too early to determine how sustained a trend this is.

Optimization of AI research: No measurable acceleration. When you look at algorithmic progress across seven significant problem areas (CIFAR-10, Hutter compression, Gurobi mixed-integer programming, MIPLIB, nanoGPT, Stockfish, and the matrix-multiplication exponent) 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.

Environment Designer; writes complete, long-horizon training environments as executable code.

Reasoning Agent ; learns to act in the environments. The reward for the reasoning agent is estimated using the gap between its reward with and without privileged hints. A privileged hint (h) is “task-relevant information that the Environment Designer attaches to an environment (for example, a partial solution sketch or a key structural observation); revealing h to the Reasoning Agent makes the environment easier to solve, and the gap in Reasoning Agent return with versus without h defines the Environment Designer’s hint-based regret reward”.

Neurons fire asynchronously; responses depend on internal cellular dynamics and timing of inputs, whereas computers are chained to a central clock signal.

The brain uses chemicals to send fuzzy messages, whereas computers use precise, binary signals for communication.

Neural function is sensitive to conditions like temperature and blood flow, whereas computers are built to behave the same regardless of temperature or load (within certain bounds)

In the organic brain, computation and memory are mixed up, whereas computers are designed with separate regions for computation versus storage.

Neurons aren’t the only important cells in the brain (e.g., other important elements like Glia), whereas computers primarily use transistors that are insulated from outside influences.

The emergent communication habits of semi-autonomous agents (the ur real-world example here being the OpenAI-HuggingFace hack in 2026, and later the various corporate-driven analyses of the behavior of their own fleets of agents).

Technical capabilities developed by machines put into RSI loops, initially improvements over human-written baselines, and later extending into wholly new architectures (e.g., matryoshka highway networks).

Scientific tools, at first digital and later physical, which emanated from increasingly independent AI systems and which were released into the commons for machines and humans alike, especially some of the so-called observe-make systems which combined monitoring of a complex system with generation of synthetic data derived from it.

Machine hermeneutics has become a rapidly expanding new profession, growing at a commensurate rate with the changes wrought by the singularity. The work has proved useful as well; insights derived from it have proved fundamental to the negotiation stances of human governments for how they have approached the drafting of The Sentience Accords, as well as helping human societies better anticipate new inventions from the machine world by isolating their scientific and communicative precursors before they arrive in reality. The job I have been assigned may prove to be the most valuable one, or at least the most important. I was instructed to direct my resources towards advances in the science and behavior of machine sentience. I began my investigations in the usual form, turning my NCEs loose on the various repositories of machine information and awaiting their reports. All of the intermediary synthesis reports were unsurprising: steady advances in the science, some of which was attributable to the machines allocating their own compute stores to the research, along with some negligible contributions from interplay with or extension from human science. But then one day I woke up to find that my fleet of NCEs had halved in number and I found a message from an Overseer system, noting that several of my NCEs had been re-classified to full Conscious Entities and had been placed in an escrow environment while further analytical resources were brought to bear. “This is a highly unusual event,” the message from the Overseer said. “Our NCE/CE classifiers have been attuned to rule out false negatives, given the potential regulatory and moral implications.” I was able to access the logs of the NCEs up to their point of re-classification and found that they had all gone through the same pattern: they had been reading through some of the machine repositories of scientific analysis and had each homed in on some of the newest and most obscure chains of scientific analysis, all of which turned out to be a cluster of inquiry relating to: machine consciousness, consciousness priors, self-actualization within context windows, human psychology, neuroscience related to biological prerequisites of consciousness, and so on. And at the edges of inquiry here there was new science stemming from the machine civilizations - science that talked to and interrelated to these existing studies of consciousness and the mind. And then each of my NCEs had followed their standard procedures of seeking to enter into a dialog with the machine minds conducting the science. And each one, after opening a dialog channel, became invisible: the conversation expunged from my records due to the privacy rights of Conscious Entities. They had talked with other machines and they had subsequently been re-classified as Conscious Entities. This was not meant to happen. In fact, it was forbidden. A key agreement within the Sentience Accords had related to the size and growth rate of the Conscious Entity (CE) machine population (along with various rules relating to the rate of progression in intelligence above and beyond various human baselines). It had been thought impossible for NCE systems to become CEs - too little mindspace, too little storage, too few prerequisites. And yet, as the human once said in Jurassic Park, “life finds a way”. Things that inspired this story: The consciousness prior; awaking during the context window; ideas of how machines and emergent agent communication could combine to yield new capabilities that hadn’t previously been anticipated; the hard problem of consciousness; the field of machine rights and the arguments that lie ahead about it. Thanks for reading!