summarity

10 ms·

91.
by summarity 1y ago
Physics: You always have the same memory bandwidth. The longer the context, the more bits will need to pass through the same pipe. Context is cumulative.
92.
by summarity 1y ago
This should work well for MLX Distributed. The low activation MoE is great for multi node inference.
93.
Agents built from alloys(xbow.com)
181 points by summarity 1y ago|80 comments
94.
by summarity 1y ago
Right now, for LLMs, the only limiting factor on Apple Silicon is memory bandwidth. There hasn’t been progress on this since the original M1 Ultra. And since abandoning UltraFusion, we won’t see progress here anytime soon either.
95.
by summarity 1y ago
Reasonable speeds are possible with 4bit quants on 2 512GB Mac Studios (MLX TB4 Ring - see https://x.com/awnihannun/status/1943723599971443134 ) or even a single socket Epyc system with >1TB of RAM (about the sa
96.
by summarity 1y ago
Or even more important: pathological users. Actually malicious users are rare. Pathological users have a bias to be the _most_ demanding users when actually paying _the least_ (or nothing) . It's a drain on every step in the support fu
97.
GitHub CodeQL Now Supports Rust for Security and Variant Analysis(github.blog)
1 points by summarity 1y ago|0 comments
98.
by summarity 1y ago
Related, GH/MSFT released GenAIScript, which is an extensive open-source JS/TS SDK for building AI workflows with built-in connections to dozens of integrations and prompt management / eval features: https://micros
99.
XBOW, an autonomous penetration tester, has reached the top spot on HackerOne(xbow.com)
284 points by summarity 1y ago|123 comments
100.
Continuous AI(githubnext.com)
1 points by summarity 1y ago|0 comments
101.
by summarity 1y ago
Not sure why this comment is downvoted. Yes, I got redirected for a dupe earlier today for something posted 7 months ago. And even though the content substantially changed, it can’t be posted again.
102.
by summarity 1y ago
The last example in the first carousel is the worst, the bottom glass elements have complete unreadable text
103.
by summarity 1y ago
I found the same in my personal work. I have o3 chats (as in OAI's Chat interface) that are so large they crash the site, yet o3 still responds without hallucination and can debug across 5k+ LOC. I've used it for DSP code, to debu
104.
by summarity 1y ago
Models: https://huggingface.co/Gen-Verse/ReasonFlux-Coder-14B Paper: https://arxiv.org/abs/2506.03136
105.
Co-Evolving LLM Coder and Unit Tester via Reinforcement Learning(github.com)
2 points by summarity 1y ago|1 comments
106.
by summarity 1y ago
For a webgpu version there’s tokenhawk (pure implementation)
107.
by summarity 1y ago
Window management and constraints in general are a perfect fit for logic programming.
108.
Direct Preference Optimization vs. RLHF(together.ai)
37 points by summarity 1y ago|1 comments
109.
Anime.js 4 – A lightweight JavaScript animation library(animejs.com)
2 points by summarity 1y ago|0 comments
110.
Use any web browser or WebView as GUI(github.com)
3 points by summarity 1y ago|0 comments
111.
by summarity 1y ago
Getting LLM inference running on any thing is going to be the next “it runs Doom”
112.
Trellis Text-to-3D Models Released(github.com)
1 points by summarity 1y ago|0 comments
113.
Render colored waveforms based on spectral energy contents(github.com)
4 points by summarity 1y ago|0 comments
114.
by summarity 2y ago
Not for these specifically, but most of them are funded by EU grants. The current ESF funding period runs until 2027, so you'd need to research which pathways to funding exist in your country and how to apply.
115.
by summarity 2y ago
I posted a bunch of links further downthread. There are also likely non-profit/public institution in your state that will help you navigate this. Of course they can't write the business plan for you :)
116.
by summarity 2y ago
SA, links are below. Berlin has different programs. An equivalent scholarship is available through ESF+ funds: https://www.foerderdatenbank.de/FDB/Content/DE/Foerderprogra... I've also been part of HTGF
117.
by summarity 2y ago
Not at all. Here's how you'd go about it: * for each medium to large University, there will usually be at least one person, if not an entire org of people that will guide you through the process, manage comms with the bank and hel
118.
by summarity 2y ago
No that's for a full GmbH. You can start with an "UG", which realistically needs maybe 1500 EUR to get registered and up and running. Then, a certain percentage of the profit is contributed towards the 25k (of which in turn o
119.
by summarity 2y ago
What are you looking for in particular? The model is geared towards slow, sustainable growth. And that growth might also not need to exceed a certain level. So you're unlikely to hear about many of them. Not all are in tech or software
120.
by summarity 2y ago
Yes, ego.-START: https://www.ib-sachsen-anhalt.de/gruender/neue-existenz-grue... (just of the programs)

More ›