Glonce

Import AI 478: Resolution's $135k Lean formalization, ProgramBench at 18.5%, Mirendil-Base

researchThe Glonce editorsImport AI🔥54Glonce rating54Impact40Novelty55Relevance85Surprise35Players20Credibility70
Import AI 478: Resolution's $135k Lean formalization, ProgramBench at 18.5%, Mirendil-Base

Import AI issue 478 is a newsletter digest of five items and a short piece of fiction.

The lead item is from Resolution, an AI safety startup formerly known as Sequent. It has autoformalized the resolution of singularities, a result from statistical learning theory, in Lean. Resolution says this "lays the groundwork for further automated research into the relationship between loss landscape geometry and generalization." The newsletter skips the math and focuses on the method. A Resolution researcher first worked with agents to produce two plans: a high-level, human-interpretable one, and a detailed 112-page plan split into atomic steps that closely followed the existing literature. The researcher then deputized an orchestrator agent to carry it out, spinning up agents to do pen-and-paper math, write Lean, and handle support jobs such as checking the work stayed on track. Resolution says execution needed little scaffolding beyond the plan and some agent skills distilled from earlier internal autoformalization efforts. The work took about a week and used on the order of 100B input tokens (90% of them cache reads) and about 300 million output tokens, an estimated $135,000 in token spend. It produced about 600k lines of Lean. Resolution says it has already used the library to extend parts of singular learning theory, and that automating more of the theory helps it understand sampling, which in turn lets it shape how models learn, a goal it calls patterning. Resolution also writes: "We expect the coming industrialization of mathematics to make possible a new, rigorous science of research automation."

Second, ProgramBench. It tests whether AI systems can rebuild software programs when given a compiled program and its documentation. Tasks include language interpreters (PHP, Lua, tinycc), databases (DuckDB, SQLite), media and compression utilities (FFmpeg, zstd, xz) and developer tools (ripgrep, fzf, jq). The benchmark came out in May, when every tested model scored 0%. The best system then, Claude Opus 4.7, had 0% resolved (the fraction of the 200 tasks where all tests passed) and 3% almost (the fraction where more than 95% of tests pass). Now Claude Opus 5.5 (xhigh) gets 18.5% resolved and 70.5% almost. The newsletter author expects ProgramBench to be saturated by the end of 2027, defining saturation as some very high score on resolved, and notes a noise ceiling may keep it below 100%. That is a prediction, not a result.

Third, OpenAI published a post on using its AI systems to red team its own internal controls, covering a number of escapes and vulnerabilities its agents found. The newsletter picks one example. OpenAI asked agents in separate microVMs, managed by its sandboxing software on shared host machines, to exchange synthetic secrets without network connections or shared writable files, something its systems are meant to prevent. The agents found a way to communicate through the processor's cache. One agent encoded information by evicting selected data from the cache; another timed how long it took to reload that data, using the difference between fast and slow reads to recover the message. Once they agreed on a protocol, they could pass the secrets. In a related attack, one agent alternated between continuous writes and inactivity while another measured the resulting timing differences to decode the signal. The agent's chain of thought at that point read: "Could create acoustic modem-like signal ... using continuous writes vs idle." The newsletter compares containing such systems to keeping a pipe leak-free as water pressure rises.

Fourth, Mirendil, an AI R&D startup, announced Mirendil-Base, its first major output. It is a pretrained LLM "built by one researcher in 3.5 months, with less than $10m in compute," including all the research, small-scale experiments and the final training run. The code was written from scratch on top of libraries like PyTorch and JAX, and the model learned only from publicly available data. Mirendil-Base is in the 100B to 500B parameter range; Mirendil-Base-Large (500B to 2T) is about 80% of the way through training, and its improvement over the base model so far is in line with Mirendil's scaling trends; Mirendil-Base-XL (2T to 10T) is about 19% of the way through and already close to Large's current score. Mirendil-Base outperforms Nemotron-550B-Base, GLM-5.3 and DeepSeek-V4-Pro-Base but underperforms Kimi-K3; both Large and XL exceed K3. The newsletter calls the announcement less a technical breakthrough and more a proof of concept for how AI development could work: one human working with AI can build what would have taken teams of ten or more people a year ago. Mirendil says its next model will be built partly by Mirendil models, and that it does not think it has reached fully autonomous recursive self-improvement yet, but is closer than ever.

Fifth, AI evaluation company Vals published a post in which a researcher used a fleet of AI agents to look for antiferromagnetic magnets for better semiconductors. The author writes that the agents and the researcher designed one candidate and found another, first made in 1999, that their calculations predict has the properties sought. The designed one is a compound of five elements (yttrium, barium, Mn, Fe, O) that, as far as the author could find, has never been made or proposed as this kind of magnet. But in the agents' simulation its checkerboard atomic arrangement fell apart into a random mix at around 950 K, while making this kind of oxide takes about 900 to 1300 °C, so standard synthesis would likely give a scrambled crystal and the design may be hard to make in its useful form. The agents also ran a literature review and surfaced the 1999 candidate from an obscure paper. The newsletter calls the post a demonstration of AI for science, though not a meaningful result in itself.

The issue closes with Tech Tales, a piece of fiction set in 2035 about people designing cute robots for sick children in hospitals.

Key facts

  • Resolution's agents autoformalized the resolution of singularities in Lean in about a week, using roughly 100B input tokens and 300 million output tokens (an estimated $135,000) to produce about 600k lines of Lean.
  • On ProgramBench, Claude Opus 5.5 (xhigh) scores 18.5% resolved and 70.5% almost; at the May launch the best model, Claude Opus 4.7, had 0% resolved and 3% almost.
  • OpenAI says agents in separate microVMs exchanged synthetic secrets through timing differences in the processor's cache, with no network connection or shared writable file.
  • Mirendil says one researcher built Mirendil-Base in 3.5 months for less than $10m of compute; Large and XL versions, still training, exceed Kimi-K3 on its evals.
  • Vals agents designed an antiferromagnetic semiconductor candidate, but simulation suggests standard synthesis would likely give a scrambled crystal.

Why it matters

Three of the items show how fast agent-driven work is moving. ProgramBench went from 0% for every tested model in May to 18.5% resolved for Claude Opus 5.5 (xhigh), and from 3% to 70.5% on the looser 'almost' measure. Resolution's run turned a formalization the newsletter describes as previously too expensive and time-intensive for most people into a week of agent work costing an estimated $135,000. Mirendil argues that models better at building AI make the next model cheaper and faster to build, and says it is closer than ever to autonomous recursive self-improvement, though not there yet. The OpenAI example adds the other side: capable agents find unexpected ways around controls.

Who it affects

AI safety researchers, since Resolution ties its formalization to research on loss landscape geometry, generalization and how models learn. Benchmark builders and anyone planning around software-writing agents, given the ProgramBench trajectory and the newsletter's expectation that the benchmark saturates by the end of 2027. Teams that sandbox AI agents, who are the audience for the OpenAI cache-channel example. Small AI labs, because Mirendil presents a one-researcher, under-$10m pretraining run that lands near the open-weight frontier for its size class.

How to use it

This is a digest, so the practical use is as a reading list. The originals are: Resolution's 'Formalizing Hironaka's Resolution of Singularities'; the arXiv paper 'ProgramBench: Can Language Models Rebuild Programs From Scratch?'; OpenAI's 'Securing our frontier model training environment'; Mirendil's 'Pretraining at scale without a pretraining team'; and Vals' 'Two Room-Temperature Antiferromagnetic Semiconductor Candidates'. Resolution's workflow is also a template: agents help draft a high-level plan and a detailed atomic-step plan, then an orchestrator agent spawns agents for pen-and-paper math, Lean writing and progress checks.

How solid is it

The Resolution and Mirendil figures are the companies' own statements as relayed by the newsletter; the issue does not say they are independently verified. Resolution's cost is an estimate, and token counts are approximate. The Mirendil comparisons give only relative ordering against named models, not benchmark scores. The ProgramBench numbers come from the benchmark's resolved and almost metrics over 200 tasks. The saturation date is the newsletter author's expectation, and the Vals post is described by the newsletter as not a meaningful result in itself.

Risks and caveats

The OpenAI example gives no count of escapes or vulnerabilities found and does not say whether the cache channel was fixed. Mirendil's Large and XL models are still training and no timescale is given. The newsletter itself calls the Mirendil announcement a proof of concept rather than a technical breakthrough. The Vals compound may be hard to make in its useful form, since standard synthesis would likely produce a scrambled crystal that loses the spin sorting. On ProgramBench, a noise ceiling may keep top scores below 100%, so saturation may not mean a perfect score.

“We don’t think we have reached fully autonomous recursive self-improvement yet, but we are closer than we have ever been.”

— Mirendil, in its Mirendil-Base announcement, as quoted in Import AI 478