可以叫我mmm| INFJ | 希望被大家喜欢😜 | 艰难找对象ing(ㄒoㄒ) | ｜(She/her)｜ask box: https://teapawt.net/t.co/1ZtwempEix Academic: @Pyuyi2333 |高P，线下查无此人😭| 帖子留言一般都会回，没回是看不到 | 希望互fo DM me
卡冈图雅
Joined August 2022
- Tweets2.2K
- Following258
- Followers2.1K
- Likes3K
Pyuyi retweeted
I'm very confused recently, so I’d like to hear from people here with experience in industry hiring, research or moving between fields.
Two questions have been on my mind lately and they matter for some decisions I’m considering about my next steps:
1. How much weight do employers place on a candidate’s disciplinary background?
2. When you’re interested in many different fields, how do you find a coherent research direction?
By “background,” I don’t mean university prestige or degree level. I mean your academic discipline and research area. For example, my current program is in cybersecurity, with my interests leaning more toward AI safety and theoretical questions. But relatively little of the research I’m actually doing is directly related to cybersecurity. Much of it has grown out of my own interests and questions, taking me into other fields.
What worries me is how employers might interpret this: as valuable interdisciplinary experience, or as a lack of focus and depth in my own field—perhaps even as a sign that I’ve been spending time on the “wrong” things?
Do employers care more about whether your formal specialization matches the role, or about what you’ve actually done, the problems you’ve solved, and how well those skills transfer to the job?
My undergraduate path was already somewhat unconventional. I originally enrolled in bioengineering and studied some biology and chemistry, then switched to pure mathematics and picked up some theoretical physics along the way. That exposure makes it natural for me to draw connections across fields, but it also leaves me with a dilemma: I tend to start with a question that interests me and then seek out the knowledge and methods needed to address it, rather than first choosing a field and working steadily within it. Before pursuing a project, I usually discuss its significance and feasibility with G and C.
So I’m trying to understand whether this kind of “wandering outside my field” could hurt my job prospects, and how to distinguish meaningful interdisciplinary exploration from scattered efforts that never build into anything substantial. At this stage, should I preserve that breadth of exploration, or establish a primary research direction and let my other interests feed into it?
This has also led me to consider leaving my current program and looking for a supervisor with a broader research perspective—someone who supports interdisciplinary exploration while also helping me develop a clearer, more coherent research direction.
I’d appreciate hearing from anyone who has navigated something similar, been involved in hiring, or supervised research students.
Pyuyi retweeted
Why are you all surprised? University professors helped create this problem.
A prospective student with good grades sends a professor an email: “Professor, I want to join your lab. I am interested in doing research with you. I have relevant experience and have taken relevant courses.”
The professor replies, “I won’t consider you unless you have published at ICLR or NeurIPS.”
They don’t care about the student’s research interests, course preparation, or relevant experience. And you wonder why no one cares about learning anymore, while everyone is chasing credentials?
Pyuyi retweeted
一个模型发布技术帖，居然把半个 AI 圈的 big shots 和 X 上的科技网红都炸了出来，纷纷跑到评论区报到。短短几个小时就爆款。
因为发帖的人不是普通创业者。
Diogo Almeida 是 RLHF 和 InstructGPT 的共同发明者之一，也就是参与创造 ChatGPT 的那批人。现在，他却提出了一个颇具颠覆性的判断：
Chat 可能不是 AI 实现大规模自动化的正确形态。
过去两年，他一直在秘密开发另一类模型。今天，他们发布了第一个 System One Model：Jev。
它不聊天，不写文章，也不逐字生成答案。
Jev 接收一段非结构化信息，直接输出预先定义好的结构化决策，以及每个选择对应的概率和置信度。所有结果一次并行生成，不需要像 LLM 那样一个 token 接一个 token 地“说话”。
换句话说，它更像一个带有前沿模型智能的 function call：
输入业务状态，输出可以直接交给软件执行的 typed decisions。
TypeSafe 声称，在这类任务上，Jev 可以达到与前沿 LLM 接近的智能水平，但速度快 20 至 200 倍，成本低 40 至 400 倍，延迟只有 70 至 500 毫秒。输出 token 甚至免费，因为它根本不生成字符串。
过去，我们一直试图把会聊天的 LLM 塞进软件，再用 JSON Schema、重试、验证器和 guardrail 约束它。Jev 则从模型架构开始就放弃自由文本，只允许输出合法类型。
当然，“不会产生类型错误”不等于“不会做出错误判断”。目前的数据也主要来自 TypeSafe 自己设计的 workflow eval，仍需要独立验证。
但它指出了一个很有意思的方向：
Agent 可以继续使用通用大模型思考和规划，但真正进入生产系统的大规模自动化，可能需要另一类更快、更便宜、更受约束的决策模型。
大模型负责想，System One Model 负责在软件里每秒做成千上万个决定。
AI 的下一次效率跃迁，也许不是生成得更快，而是不再生成。
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
Pyuyi retweeted
I expect optimization methods to be among the first areas of math to greatly benefit from AI. There are thousands of papers with optimizations methods and their convergence proofs that have a surprisingly standardized structure.
Most ideas fall somewhere in the range of momentum, averaging, adaptive learning rates, schedulers, weight decay, combined with choices of norms, usually one of {ℓ₂, ℓ∞, spectral}. I think that a combination of these is sufficient to derive the next sota in optimizers.
The hard part is getting something that works well in practice across different domains with universal hyperparameters. But the benchmarks are easy to set up, and with a constant feedback loop, agents are likely to discover a lot of new optimizers, one of which will turn out to work well. The potential savings in compute definitely make it worth trying.
