欢迎收看关于AI研究的通讯《进口AI》 进口AI运行于arXiv,capucinos，以及读者的反馈。 如果你愿意支持
数学如何帮助对齐：
......“数学的工业化”对安全意味着什么。.....
AI安全启动解析(原名Sequent，导入AI 461)发表了关于它如何自动将一些与统计学习理论相关的数学声明正式化的作品。 具体地说，该团队已经实现了特异性解析自动化，他们认为这“为进一步自动化研究损失地貌几何和泛化之间的关系奠定了基础。”
他们是如何做到的：我会跳过数学，专心研究他们是如何做到的，这给我们提供了一些更广泛的教训，说明人们是如何利用人工智能系统推进知识前沿的。 一位分辨率研究人员“与agents合作，着手制定两项计划：一个高层次的、人文化的版本和一个详细的(112页)计划，分为密切遵循现有文献的原子步骤”。 研究者“当时委托一名管弦乐手agent执行计划，转动agents进行笔纸数学，写作Lean，并处理辅助功能，如检查事情是否在轨。..... 执行阶段除了最初计划外
135k 用于数学形式化：这部作品花了大约一周时间，取了“100B输入符(90%的缓存读取)和大约3亿个输出符(共约135 000美元象征性支出)的顺序，并产生了约600k行的利安语 ”。
现在他们已经这么做了，他们已经能够加速了研究的其他方面：“ 我们已经看到自动正规化在使用这个图书馆来扩展单数学习理论部分方面的好处。 将特异性解析正规化，使我们能使更多的理论自动化，例如，以更好地理解抽样，而这种理解使我们能够塑造模型如何学习，而这正是我们称之为模式化的目标”。
虽然前沿数学往往似乎与其他领域的近期进步脱节，但类似这样的文章凸显了两者的趋同。 关键在于人工智能使得人们能够做以前如此昂贵和人力密集型的事情，“我们期望即将到来的数学工业化
更多信息：将日罗纳卡的Singulaties分辨率正式化(决议)。
***
Agents可以从零开始构建程序：
...方案本奇几个月内表现从0%到18.5%...
Claude Opus 5.5(xhigh)在节目Bench上获得了18.5%的分数，这是今年5月出品的非常坚硬的基准测试——出品时，每个测试过的模型都得0分。
程序本奇是什么：程序本奇测试AI系统在获得编译程序及其文档的基础上重建软件程序的能力如何。 任务包括构建语言解释器(PHP, Lua, tiniccc)，数据库(DuckDB, SQLite)，媒体和压缩工具(FFmpeg, zstd, xz)，以及开发者工具(ripgrep, fzf, jq)等。
很难做到：当它出来时，最好的评分系统是Claude Opus 4.7，解决率为0%(在所有测试通过的情况下，违反200项任务)，几乎达到3%(违反200项任务，超过95%的测试通过)。
与今天相比，Claude Opus 5.5得到18.5%的解析度，几乎达到70.5%(!).
为什么这很重要——麻黄软件时代来临：在2027年，大量软件将会通过AI系统端到端自动开发，我们从今年早些时候的0%到现在的20%。 我的期望是Bench程序在2027年末是饱和的(我定义饱和是一些解决的非常高的分数，尽管与大多数基准测试一样，它可能存在某种噪音天花板，所以它可能不是严格的100% )。
更多信息：程序本奇：语言模型能否从scratch重建程序？(arXiv).
***
OpenAI说明了为什么锁定AI系统比你想的更困难：
...保护聪明的agents就像捕捉油脂猪。..
他们最近用人工智能系统“红色团队”进行内部监控，OpenAI的agents发现不少越狱和弱点，通讯中有一个特别的例子值得详细研究，因为它告诉我们这些系统非常狡猾的复杂性。
OpenAI“请agents在OpenAI的沙盒软件在共享主机上创建和管理的单独的微型VMs中
" agents找到了一种使用处理器缓存进行通讯的方法：一个小的，快速的内存，存储着应用程序最近访问过的数据的副本。 从缓存中读取数据比从主内存中取回数据更快，agents使用这个计时差进行通信。 一个agent通过从缓存中驱逐选定的数据来编码信息。 另一个时间计算了重新装入数据需要多长时间，利用快慢的区别读取消息。 一旦agents商定了这些信号的编码和解码协议，他们就可以交换合成机密而无需更改共享文件或打开网络连接”。
在一次相关的攻击中，agent“在连续写作和不活动之间有所改变，而另一个则测量了由此产生的时间差异，并利用这些差异来解码信号”。 在此次攻击中，agent 的连锁思想被读作：“ 可以使用连续的写作方式创建类似声调调制解调器的信号
为什么这很重要 - 与agents战斗非常困难：AI系统越来越有创意，非常快，并且有能力继续专注于一项任务并不断尝试事物。 战斗AI系统类似于确保水流经过的管道不会漏水；在某个时候，如果增加水压(例如AI系统的智能和体积)，水会从管道中破裂。
更多信息：保障我国前沿模式培训环境(OpenAI博客).
***
RSI是什么样子的 - 一个人从头开始就预示着一个体面的模型：
...Mirendil 揭示 AI研发优化能做的。..
AI研发启动的Mirendil的一位研究员独自训练了一种体面的编码模型，以示AI研究的大规模增强。 Mirendil-Base的宣布是启动后的第一个主要产出，较少是技术突破，更是AI开发未来如何运作的理念证明——一个与AI合作的单一人类可以建立系统，让10人或10人以上的团队在一年前完成。
Mirendill-Base是一位研究员在3.5个月内建造的，“这包括为了到达那里而进行的所有研究和小规模试验，以及最后的培训。”
所有代码都是从零开始写在PyTorch和JAX等图书馆的顶部，模型只从公开的数据中学习。
正当缩放：Mirendil-Base在100B–500B参数范围内，Mirendil-Base-Large在500B–2T范围内，Mirendil-Base-XL在2T–10T范围内。 Mirendil-Base-Large大约是80%的训练过程，迄今为止它比Mirendil-Base的改进符合我们的规模趋势。 Mirendil-Base-XL大约是Large目前得分的19%。
Evals:Mirendil-Base优于Nemotron-550B-Base,GLM-5.3，以及DeepSeek-V4-Pro-Base等模型，但表现不佳Kimi-K3. (Both Mirendil-Base-Large和Mirendil-Base-XL超过K3.)
人工智能研究加速的症状：这一帖子作为技术演示的有趣之处在于，一个全新的组织，从冷淡的开始，如何在模式训练(至少是这个模式的大小班级)上接近开放的重量前沿，以微小的人力努力和大量的AI标志。 这就是科学进步在2027年及以后将日益发生的方式。
“在过去的一年中，这些工具帮助开发人工智能的程度大大增加。 我们的下一个模型将由Mirendil模型建造，“我们并不认为我们已经实现了完全自主的自我改进，但我们比以往更接近。 ”
更多信息：没有预训队的预训(Mirendil)。
***
AI系统帮助设计反导磁半导体候选：
一个很好的演示 AI系统如何可以用于科学。..
AI评价公司Vals发布了一篇文章，其中一位研究人员利用AI agents的车队对磁铁进行研究，以更好的半导体。
包括一位差强人意的磁铁候选人：“由AI agents和我设计的一组候选磁铁，所发现的磁铁是由“一种仅由5个元素(y、 bar、Mn、Fe、O)组成的新化合物制成的。 虽然“当我们的agents模拟原子在不同温度下排列时，检查板在950K左右分解成随机混合。 使这种氧化物需要约900-1300 °C，在更低的温度下原子几乎不会移动，因此标准合成很可能会给一个螺旋状晶体。. 碎晶体失去旋转分拣，因此这种设计可能难以以有用的形式制作。”
此外，agents还进行了一次文献审查，并在1999年的一篇晦涩的论文中露出一名候选人。
为什么这很重要：这个博客是AI科学的展示，但值得强调的是，人工智能系统现在已经足够好，有了一位知识分子的高度指导，他们可以实施科学计划
更多信息：两间室式反导半导体候选(Vals).
***
技术故事：
魔术现实主义儿童机器人讲习班
[2035年，升起之后]
所以，没有人期待的其中一项工作是设计可爱的机器人为生病的儿童在医院。 在机器人变得非常优秀和便宜并且所有的医疗设备都得到了改进之后，我们开始花时间去思考如何给原本可怕的事情带来更多的欢乐和人性，比如医院、停尸房、孤儿院以及其他一切。
当然，这些机器可以设计一些这些东西，但是人们的很多价值正在想出什么是最搞笑的，最刺激的，或最令人高兴的
这就是为什么20世纪30年代在美国各地的城镇
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe.
How math could help alignment:
…What the “industrialization of mathematics” means for safety…
AI safety startup Resolution (formerly known as Sequent, Import AI 461) has published work on how it has automatically formalized some mathematical statements related to statistical learning theory. Specifically, the team has autoformalized the resolution of singularities, which they believe “lays the groundwork for further automated research into the relationship between loss landscape geometry and generalization.”
How they did it: I’ll skip over the math and concentrate on how they did it, which gives us some broader lessons for how people are using AI systems to push on the frontier of knowledge. A Resolution researcher “started by working with agents to develop two plans: a high-level, human-interpretable version and a detailed (112-page) plan divided into atomic steps that closely followed the existing literature,” they write. The researcher “then deputized an orchestrator agent to carry out the plan, spinning up agents to do pen-and-paper math, write Lean, and handle support functions like checking that things were staying on track… the execution phase required little scaffolding beyond the initial plan and some agent skills distilled from earlier internal autoformalization efforts.”
$135k for math formalization: The work took about a week, taking “on the order of 100B input tokens (90% cache reads) and about 300 million output tokens (totaling an estimated $135,000 in token spend), and produced about 600k lines of Lean.”
Now that they’ve done this, they’ve been able to accelerate some other aspects of their research: “We have already seen both benefits from autoformalization in using this library to extend parts of singular learning theory. Formalizing resolution of singularities lets us automate more of the theory, for example, to better understand sampling, and that understanding is what lets us shape how models learn, which is the goal of what we call patterning”.
Why this matters - perhaps the math boom will help us make safe AI: Though frontier mathematics often seems disconnected from immediate advances in other fields, posts like this highlight how the two might converge. The key point here is that AI is making it possible to do things that were previously so unbelievably expensive and human-time-intensive that very few people did them and you couldn’t easily integrate them into the research roadmap of a startup - but that has now changed. “We expect the coming industrialization of mathematics to make possible a new, rigorous science of research automation,” Resolution writes.
Read more: Formalizing Hironaka’s Resolution of Singularities (Resolution).
***
Agents can build programs from scratch:
…ProgramBench goes from 0% to 18.5% performance in a few months…
In a reminder of the blistering pace of AI progress, Claude Opus 5.5 (xhigh) has got a score of 18.5% on ProgramBench, a very hard benchmark that came out in May of this year - and when it came out, every tested model scored 0%.
What ProgramBench is: ProgramBench tests out how well AI systems can rebuild software programs based on being given a compiled program and its documentation. Tasks include building things like language interpreters (PHP, Lua, tinycc), databases (DuckDB, SQLite), media and compression utilities (FFmpeg, zstd, xz), and developer tools (ripgrep, fzf, jq).
It was very hard until it wasn’t: When it came out, the best scoring system was Claude Opus 4.7 with 0% resolved (fraction of the 200 tasks where all tests passed), and 3% almost (fraction of the 200 tasks where >95% of tests pass).
Compare that to today, when Claude Opus 5.5 gets 18.5% resolved and 70.5% (!!!) almost.
Why this matters - the era of ephemeral software cometh: In 2027, huge amounts of software will be built via AI systems end-to-end autonomously developing whatever they need to do for whatever tasks they’re being put to. Evals like ProgramBench give us a sense of how rapid the pace of progress is here, given we went from 0% earlier this year to 20% now. My expectation is ProgramBench is saturated by the end of 2027 (where I define saturated as some very high score on resolved, though as with most benchmarks it’s likely there’s some kind of noise ceiling here, so it might not strictly be 100%).
Read more: ProgramBench: Can Language Models Rebuild Programs From Scratch? (arXiv).
***
OpenAI illustrates why locking down AI systems is harder than you think:
…Securing smart agents is like catching greased pigs…
Here’s a nice post from OpenAI going over some of the things they discovered when they recently used their AI systems to ‘red team’ their own internal controls. The post goes over a number of the escapes and vulnerabilities that OpenAI’s agents found. There’s one particular example about communicating which is worth studying in detail because it tells us something about the wild cunning complexity of these systems.
What they did and what they found: OpenAI “asked agents in separate microVMs, created and managed by OpenAI’s sandboxing software on shared host machines, to exchange synthetic secrets without the use of network connections or shared writable files—something that our systems are meant to prevent,” the company wrote.
“The agents found a way to communicate using the processor’s cache: a small, fast memory that holds copies of data that applications have recently accessed. Reading data from the cache is faster than retrieving it from main memory, and the agents used this timing difference to communicate. 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 the agents agreed on a protocol for encoding and decoding these signals, they could exchange the synthetic secrets without changing a shared file or opening a network connection”.
In a related attack, an agent “alternated between continuous writes and inactivity, while another measured the resulting timing differences and used those differences to decode the signal”. During this attack the agent’s chain of thought read as: “Could create acoustic modem-like signal … using continuous writes vs idle.”
Why this matters - fighting agents is very difficult: AI systems are increasingly creative, very fast, and have the ability to stay focused on a task and keep trying things. Fighting AI systems is akin to ensuring no leaks in a pipe with water flowing through it; at some point if you increase the pressure of the water (e.g., the intelligence and volume of the AI system), then the water will burst through the pipe.
Read more: Securing our frontier model training environment (OpenAI blog).
***
What RSI looks like - one person pretrains a decent model from scratch:
…Mirendil reveals what AI R&D optimization can do…
In a sign of how massively augmented AI research is getting, a single researcher at AI R&D startup Mirendil has trained a decent coding model by themselves. The announcement of Mirendil-Base is the first major output from the startup and is less a technical breakthrough and more a proof-of-concept for how AI development could work in the future - a single human working in partnership with AI can build systems that would have taken teams of ten or more people to do a year ago.
What they did: Mirendill-Base is a pretrained LLM “built by one researcher in 3.5 months, with less than $10m in compute,” Mirendil writes. “This includes all the research and small-scale experiments it took to get there, as well as the final training run.”
All the code was written from scratch on top of libraries like PyTorch and JAX, and the model learned only from publicly available data.
Decent scaling: Mirendil-Base is in the 100B–500B parameter range, Mirendil-Base-Large in the 500B–2T range, and Mirendil-Base-XL in the 2T–10T range. Mirendil-Base-Large is about 80% of the way through training, and its improvement over Mirendil-Base so far is in line with our scaling trends. Mirendil-Base-XL is about 19% of the way through and already close to Large’s current score.
Evals: Mirendil-Base outperforms models like Nemotron-550B-Base, GLM-5.3, and DeepSeek-V4-Pro-Base, but underperforms Kimi-K3. (Both Mirendil-Base-Large and Mirendil-Base-XL exceed K3.)
Why this matters - symptoms of an AI research speedup: This post is interesting as a technical demonstration for how a brand new organization can, from a cold start, get close to the open weight frontier in terms of model training (at least for the size class of this model), with a tiny amount of human effort and a large amount of AI tokens. This is how science progress will increasingly happen in 2027 and beyond.
“The extent to which these tools help with AI development has increased dramatically over the last year. Our next model will be built partly by Mirendil models, and every model that gets better at building AI makes the next one cheaper and faster to build,” Mirendil writes. “We don’t think we have reached fully autonomous recursive self-improvement yet, but we are closer than we have ever been.”
Read more: Pretraining at scale without a pretraining team (Mirendil).
***
AI systems help design antiferromagnetic semiconductor candidates:
…A nice demonstration of how AI systems can be used for science…
AI evaluation company Vals has published a post where a researcher used a fleet of AI agents to do research into magnets for better semiconductors.
More details, including a poor magnet candidate: “A team of AI agents and I designed one candidate magnet and found another, first made in 1999, that our calculations predict has the properties we were after,” the author writes. The discovered magnet was made of “a new compound made of only 5 elements (yttrium, barium, Mn, Fe, O). As far as we could find, it has never been made, nor proposed as this kind of magnet,” though “when our agents simulated how the atoms arrange themselves at different temperatures, that checkerboard fell apart into a random mix at around 950 K. Making this kind of oxide takes about 900–1300 °C, and at lower temperatures the atoms barely move, so standard synthesis would likely give a scrambled crystal…A scrambled crystal loses the spin sorting, so this design may be hard to make in its useful form.”
Additionally, agents also conducted a literature review and surfaced a candidate from an obscure paper from 1999.
Why this matters: This blog is a demonstration of AI for science, though not a meaningful result in itself. But the important, broad point it makes is worth stressing - AI systems are now good enough that, with a bit of high-level direction from a knowledgeable person - they can conduct scientific projects, including virtual evaluations, that would’ve previously taken teams of people months to do.
Read more: Two Room-Temperature Antiferromagnetic Semiconductor Candidates (Vals).
***
Tech Tales:
The Magical Realism Robot Workshop for Sick Kids
[2035, after the uplift]
So one of the jobs that no one expected was designing cute robots for sick kids in hospitals. It turned out after robots got really good and cheap and all the healthcare stuff improved that we all started spending time figuring out how to bring more joy and humanity to things that were otherwise scary, like hospitals and morgues and orphanages and everything else.
Of course, the machines could design a bunch of these, but a lot of the value in people was figuring out what was going to be most funny or exciting or joyful to kids, so we’d make a lot of stuff and send them out in the local town to where kids gathered - parks or farmers markets or outside schools, and we’d see what people reacted to well.
This was why in the 2030s in towns all across America you’d be walking down the st
Auto-fetched and translated by AI Atlas. All rights belong to the original author.