ทวีตที่ปักหมุด
we're on track to building the best dspy powered llm framework in the typescript world
Omar Khattab@lateinteraction
@VarunGodbole @isaacbmiller1 @DSPyOSS Yes — @dosco’s Ax is pretty much the official DSPy in TS.
English
spacy
16.2K posts
spacy
@dosco
LLM research, systems and compilers | ax + dspy in TS | agent engineering
เข้าร่วม Temmuz 2008
1.6K กำลังติดตาม4.9K ผู้ติดตาม
spacy รีทวีตแล้ว
guess what NVIDIA used here for an "attention-based encoder-decoder to retrieve directly from its own internal representations"?
late interaction is sparse attention
Sumit@_reachsumit
Retrieval from Within: An Intrinsic Capability of Attention-Based Models NVIDIA enables encoder-decoder models to perform retrieval directly through their own cross-attention mechanism, eliminating the need for a separate retriever. 📝 arxiv.org/abs/2605.05806
English
spacy รีทวีตแล้ว
Work led by @jacobcares showed that little compute for building an LLM is actually in the final runs. The vast majority of compute goes to developing a recipe.
Creating the recipe openly is a huge lever in making sure the research community's compute pushes to new knowledge.
Ai2@allen_ai
Today we’re bringing new NSF OMAI compute online with NVIDIA Blackwell Ultra-powered systems, turning a $152M national investment from @NSF & @NVIDIA into a foundation for truly open AI research. 🧵
English
as the models go bigger and the RL environments far more complex spanning every little task on the planet you really need to think in exponentials there is no option but space, expect every lab come around to this very soon
Tom Brown@nottombrown
Terrestrial datacenters will increasingly be bottlenecked by permitted real estate space. Lots of space in space.
English
list of shit canada doesn't need
- summits
- endless debate
- innovation centers
- government intervention
- government purchased ai data centers
- tech conferences sponsored by boomer companies
things we need
- cracked founders maximizing shareholder value
English
looking at my own personal inference demand and the rate at with it's growing. rate limit anxiety is a real thing
Daniel A. Saedi (DataManDan)@TheRealDanSaedi
Last week Jeff Dean and several other notable and experienced computer legends got together for dinner and came to the conclusion that demand for AI inference in 2030 will be 10000x higher than today If that’s the way we’re heading then nothing else matters other than AI
English
i find the idea of ambient agents working for you fascinating. this is my goal with the dspy+rlm agent i'm experimenting with. and i obviously want to go crazy with those tokens so smaller models running locally
Ishan@radshaan
Claude now sends me random artifacts to read straight to my kindle
English
reading this book will leave you with a strong urge to run your own local agents
deepa 🏴☠️@deeparocks
i can’t tell you how much im enjoying reading accelerando a window into the future where money doesn’t matter
English
The Internet disaggregated information. AI reaggregates it.
English
golang is very very fast
spacy@dosco
tachyon was for a research paper that LLMs can vibe code complex infra and be correct and fast it. a side effect of this was that it was faster than nginx and pingora (c and rust) while being written in golang. also its 100% self contained (zero dependencies) and supports http 1 and 2. benchmarks are complex but heres some of the ones we did. github.com/dosco/tachyon/…
English
tachyon was for a research paper that LLMs can vibe code complex infra and be correct and fast it. a side effect of this was that it was faster than nginx and pingora (c and rust) while being written in golang. also its 100% self contained (zero dependencies) and supports http 1 and 2. benchmarks are complex but heres some of the ones we did. github.com/dosco/tachyon/…
English
@dosco @caddyserver Well, there's also HAProxy (written in C by the way).. And in our benchmarks it's very fast.
Here's a relevant benchmark video (by @antonvputra):
youtube.com/watch?v=h-ygQb…
YouTube
English
Web servers written in Go are immune to this entire class of vulnerabilities.
Cyber Security News@The_Cyber_News
⚠️ Critical Apache HTTP Server Flaw Exposes Millions of Servers to RCE Attacks Source: cybersecuritynews.com/apache-http-se… The Apache Software Foundation has released a critical security update for Apache HTTP Server, patching five vulnerabilities, including a dangerous double-free flaw capable of enabling Remote Code Execution (RCE) in version 2.4.67, released on May 4, 2026. All users running version 2.4.66 or earlier are strongly urged to upgrade immediately. The most severe of the five vulnerabilities is CVE-2026-23918, rated High with a CVSS base score of 8.8. The flaw is a double-free memory corruption bug triggered within Apache's HTTP/2 protocol implementation during an "early stream reset" sequence. #cybersecuritynews #vulnerability
English
@ferron_web @caddyserver i benchmarked against cloudflare pingora and nginx and it’s faster or the same
English
English