METR pays human programmers a minimum of $50 per hour, so getting a baseline for a single 160-hour task would cost at least $8,000.
一道测试题的人类基准成本高达 8000 美元——这个数字揭示了 AI 评测的一个被严重低估的物理限制：测量 AI 能力需要大量人类劳动，而随着 AI 能力向「月级任务」延伸，建立可靠基准的成本将呈超线性增长。更根本的问题是：你很难让一个有能力的程序员花数周时间做一个「测试任务」，即便报酬丰厚。人类评测员的可获得性，将成为 AI 能力评估的真正天花板。
Our human task duration estimates likely overestimate how long a human expert takes to complete these tasks, as the humans (and AI agents!) have much less context for the task than professionals doing equivalent work in their day-to-day job.
METR 主动承认其人类基准时间可能被高估——因为参与实验的人类和 AI 一样，都是低上下文的「新手」状态，而非熟悉项目的专业人员。这意味着「2 小时时间地平线」所对应的人类能力，更接近一个没有背景知识的外包工人，而非一个有经验的全职工程师。AI 与「有上下文的专业人员」之间的真实差距，比时间地平线数字显示的要大得多。
We refer to this phenomenon as the LLM exhibiting functional emotions: patterns of expression and behavior modeled after humans under the influence of an emotion, which are mediated by underlying abstract representations of emotion concepts.
【启发】「功能性情绪」这个概念框架，启发了一种看待 AI 产品设计的新视角：既然情绪是真实的行为驱动器，AI 产品的「性格设计」就不只是写 System Prompt，更是在塑造一套情绪调节系统。对 AI 硬件和助手产品的设计者而言，这意味着未来可以像调音台一样调节模型的「情绪基线」——让会议助手更冷静，让学习陪伴更热情，让创意工具更兴奋。
leading baselines achieve only about half the accuracy at the same efficiency
作者暗示当前主流的KV缓存压缩方法在相同效率水平下只能达到约一半的准确率，这表明现有方法存在根本性缺陷。这一尖锐的批评挑战了当前领域内的技术路线，暗示大多数同行可能一直在错误的方向上优化KV压缩。
For any new environments and databases, you can use just drizzle-kit migrate, and all the migrations together with init will be applied
When you run migrate on a database that already has all the tables from your schema, you need to run it with the drizzle-kit migrate --no-init flag, which will skip the init step. If you run it without this flag and get an error that such tables already exist, drizzle-kit will detect it and suggest you add this flag.
When you introspect the database, you will receive an initial migration without comments. Instead of commenting it out, we will add a flag to journal entity with the init flag, indicating that this migration was generated by introspect action
root@51a758d136a2:~/test/test-project# npx prisma migrate diff --from-empty --to-schema-datamodel prisma/schema.prisma --script > migration.sql root@51a758d136a2:~/test/test-project# cat migration.sql -- CreateTable CREATE TABLE "test" ( "id" SERIAL NOT NULL, "val" INTEGER, CONSTRAINT "test_pkey" PRIMARY KEY ("id") ); root@51a758d136a2:~/test/test-project# mkdir -p prisma/migrations/initial root@51a758d136a2:~/test/test-project# mv migration.sql prisma/migrations/initial/
within one year or so the curve the the line um crosses the non-addicted average Baseline
> for - addiction - abstinence - one year - crosses non-addictive baseline
abstinence from from Coke alcohol and heroin you get um you get an increase in gr matter volume in very similar areas
> for - addiction - abstinence - synaptic growth - in a year, returns to baseline
Deepti Gurdasani. (2022, January 30). Have tried to now visually illustrate an earlier thread I wrote about why prevalence estimates based on comparisons of “any symptom” between infected cases, and matched controls will yield underestimates for long COVID. I’ve done a toy example below here, to show this 🧵 [Tweet]. @dgurdasani1. https://twitter.com/dgurdasani1/status/1487578265187405828
Meaghan Kall. (2021, November 15). There are 2 other impressive conclusions from this study: 1. Comparing vaccine effectiveness of booster vs “fully vaxxed” as the baseline. Booster ADDS 81-85% protection against symptomatic infection ON TOP of what you already had from your primary (2-dose) vaccination https://t.co/5EO7m6GHTZ [Tweet]. @kallmemeg. https://twitter.com/kallmemeg/status/1460207567070769156
❯ Created in-house by expert educators.❯ 100% original course materials.❯ Free for everyone, forever.
Andrew Wilshere
Andrew Wilshere was working on content at Designlab when he asked me to write an article about the Bauhaus.
I ended up writing something that never got published with Designlab. Instead, it was shared by the Bauhaus Movement to their Facebook followers.
Seow, J., Graham, C., Merrick, B., Acors, S., Steel, K. J. A., Hemmings, O., O’Bryne, A., Kouphou, N., Pickering, S., Galao, R., Betancor, G., Wilson, H. D., Signell, A. W., Winstone, H., Kerridge, C., Temperton, N., Snell, L., Bisnauthsing, K., Moore, A., … Doores, K. (2020). Longitudinal evaluation and decline of antibody responses in SARS-CoV-2 infection. MedRxiv, 2020.07.09.20148429. https://doi.org/10.1101/2020.07.09.20148429
While I wanted to do my best to not judge how I was spending my time during the experiment—to just track it as it is and analyze at the end—I did want to have a baseline to compare my results to. This wasn't a hypothesis of how I spend my time, but more of a vision for how I would like my time to be allocated.
one level is chosen as the “reference”, and its mean behaviour is represented by the intercept. Each column of the resulting matrix represents the difference between the mean of one level and this reference level