Jasper Lu

@lu__jasper

continually learning. formerly taught models to design @figma, to see @nuro

NYC

Joined July 2009

Replies Reposts Media

Pinned
Jasper Lu
@lu__jasper
Sep 10
This is notable. DeepSeek, a lab usually first to pioneer novel algorithms and architectures, is saying that at this point, the ROI of improving data quality far exceeds that of working on novel post-training algorithms. I think this has already been true for some time for
DeepSeek
@deepseek_ai
Sep 10
🚀 Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient. 🔹 Introducing the smallest model in our new architecture family, with native visual understanding. 🔹 Designed for greater capability, faster inference, higher throughput, and scaling to larger models. 1/6
57
Jasper Lu
@lu__jasper
8h
I’m going to be stubborn here and say that the future of agent harnesses is still in cli / repls. MCPs are great, but until the norm is to wrap and call them as functions in a cli or repl, they are fundamentally held back. The reason here is composability and state management,
Thariq
@trq212
Sep 15
I was not expecting things to go this way, but I think MCPs are better than CLIs for most integrations. The models have gotten much better at tool calling, we can defer tools & MCP is now stateless. If you need to compose/filter data, add params like query to your MCP tools.
Jasper Lu
@lu__jasper
10h
what's old is new
Xudong Han
@Xudong07452910
21h
真实实验太贵怎么办？先让 Agent 在自己的历史轨迹里「做梦」几百遍，再决定下一步把算力花在哪里。这篇 Dream-RSI 做的就是这件事。每次真实探索后，系统会把走过的分支、失败方案、评测结果和计算成本保存成一棵 Discovery Tree。这些历史随后被重组成一个 Replay Simulator。Agent
Jasper Lu
@lu__jasper
19h
Big fan of these two use cases the team is targeting. I’ve thought for a while that high-recall search will eventually go the way of map-reduce. Today’s models just make this prohibitively expensive.
Diogo Almeida
@CompleteSkeptic
Sep 15
02:56
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
Jasper Lu
@lu__jasper
Sep 15
I worked in self-driving cars for quite some time, and computer use agents feel kind of analogous to me. If we could redesign the world from scratch, full AV deployment would be achievable within just a few years. We could build up completely different infrastructure, force V2X
Hang Huang
@hanghuang_
Sep 14
I might be wrong, but I am not bullish on computer use. The models are insanely smart. But we are forcing them to watch a screen, find a button, click it, wait, and check the result. All through an interface designed for humans. The token economics and speed are just not good.
2