本周的公开讨论涵盖了多个AI领域的重要信息。首先，ACX Grants的申请截止日期临近，鼓励符合条件的申请者抓紧时间。其次，Anthropic正在招聘研究工程师加入其模型福利团队，专注于AI的意识和情感研究，并欢迎有相关背景的国际人才。此外，UK AISI将发放1500万英镑的AI对齐研究资助，支持不同规模的项目，并列出了合作机构和申请流程。文章还澄清了关于Orchid公司阿尔茨海默病预测器方法论的争议，详细解释了其统计方法的应用。最后，新发布的订阅者专属内容是一篇关于“创伤”的梦境书评，尽管内容虚构，但也展现了独特的创作视角。

**AI Grants与招聘机会：** ACX Grants的申请即将截止，为AI研究者提供了重要的资助机会。同时，Anthropic正在积极招聘研究工程师，特别是针对模型福利团队，该职位对ML、哲学、神经科学或认知科学背景的专业人士开放，提供高薪并支持国际申请者。

**AI对齐研究资助：** UK AISI宣布将投入1500万英镑用于AI对齐研究项目，支持从早期种子轮到较大规模的项目，合作机构包括Anthropic和DeepMind等顶尖AI公司，申请者可关注其优先领域并按时提交申请。

**AI预测器方法论澄清：** Orchid公司就其阿尔茨海默病预测器的统计方法与Herasight的批评进行了回应，详细解释了其风险分层方法中优势比（Odds Ratio）的计算方式，以及在与文献比较时为保证公平性所使用的评估指标（AUC）和协变量。

**订阅者专属内容：** ACX发布了一篇关于“创伤”的书评，作者描述了一个关于撰写书评的生动梦境，并据此创作了内容。尽管书和部分内容是虚构的，但这种独特的创作方式为读者提供了新颖的视角。

This is the weekly visible open thread. Post about anything you want, ask random questions, whatever. ACX has an unofficial subreddit, Discord, and bulletin board, and in-person meetups around the world. Most content is free, some is subscriber only; you can subscribe here. Also:

1: This is your last chance to apply for this year’s ACX Grants. Deadline is end-of-day PST this Friday.

2: Anthropic is hiring a research engineer for the Model Welfare team - ie figuring out whether their AIs are conscious or have feelings or something, and if so how to make sure they’re okay. Candidates should have expertise in ML and maybe philosophy/neuroscience/cogsci. Job is office-remote hybrid with the office in SF, salary is $315K+, non-Americans are welcome to apply and see if Anthropic can sponsor their visa. Learn more / apply here.

3: UK AISI is looking to distribute £15m in AI alignment funding, for projects that need anywhere from a $100K pre-seed up to $1-2m. Collaborators included Anthropic, DeepMind, etc. See their priority areas and apply here by September 10th.

4: In the post on embryo selection, I mentioned that Herasight criticized Orchid's Alzheimer's predictor. A representative of Orchid reached out to say they stood by their methodology:

Herasight seems to be misreading our whitepaper. The “Performant Alzheimer’s disease risk stratification” section is meant to show the kind of performance patients can expect—people in the top 5% have an OR of 5.80, top 3% is 7.35, and top 1% is 11.69. This odds ratio is what is used to present embryo disease risk to patients and does not include covariates. The “Comparison to Published Benchmarks” section is just about comparing our models to others in the literature. To allow a head to head comparison, we used the same metric (AUC) and covariates as the paper we’re comparing against. However, to avoid future confusion, we’ve just added a sentence clarifying the AUC without covariates (0.724).

They also state that Herasight, like themselves, has only validated the predictors where there’s enough data (e.g. not schizophrenia), and they object to Herasight claiming superiority in this area.

5: New subscribers-only post - Dream Book Review: The Deal With Trauma. “Last week, I had an unusually vivid dream about writing a book review for ACX. When I woke up, I remembered the review almost word-for-word. In some sense this is a best case scenario - write posts in my sleep, and spend my waking hours relaxing on the beach - but unfortunately the book I was reviewing doesn’t exist and most of what I say about it doesn’t make sense. Still, I’m posting [it] here.”