ubermon

6 ms·

31.
by ubermon 2mo ago
it is indeed surprising how much a small model can do! As you said, for specific tasks like coding, the training data quality and new architecture of newer models probably beats model size
32.
by ubermon 2mo ago
yes. I think there is a threshold of some sort where if a model pass it can tolerate less system prompts. For example, same short prompt could work with deepseek v4 flash but not so much with Qwen 3.6 27B.
33.
Claude Code Cut Their System Prompt by 80%. Does That Work for Small Models Too?(antigma.ai)
7 points by ubermon 2mo ago|4 comments
34.
by ubermon 2mo ago
yeah, I think per user turn makes sense. But I am more inclining to only include it at session start.
35.
by ubermon 2mo ago
If you are interested in truly battle tested one, check this https://ante.run/start/philosophy
36.
by ubermon 3mo ago
For some frontier models like Fable 5 it doesn't matter, but for models less trained on long horizon tasks it very useful.
37.
by ubermon 3mo ago
I think it is more on distrusting python/typescript in general. Those dynamic scripting language are too easy to poison.
38.
by ubermon 3mo ago
we need a real agent that represent our own interest and deal with a more powerful general agents. And we can delegate answering this question to those more personal agents.
39.
by ubermon 3mo ago
lol same here, this is what I use https://github.com/AntigmaLabs/ante-preview
40.
by ubermon 3mo ago
I think this is mostly being overengineered. At this stage, a curated MEMORY.md is enough: short, readable at session start, and easy to correct when it becomes stale. The session history is another database like any other database, can be
41.
by ubermon 3mo ago
> Both scores come from a fixed fallback rule: Opus 4.8 and Sol xhigh run first; games scoring below 80 are rerun with Fable 5 and Sol max, respectively, and the higher per-game score is retained. hmm, this is like pass@n until you get t
42.
by ubermon 5mo ago
Thanks!
43.
by ubermon 5mo ago
Totally agree! I think we are very early on discovering the full potential of local models
44.
by ubermon 5mo ago
Thank you! I think there is a lot to dive deep later with different hardware, inference engine, prompt/harness setup etc.
45.
by ubermon 5mo ago
Not yet, will conduct a more comprehensive one later
46.
Open-weight 27B hits 38% on Terminal-Bench 2.0 (Opus 4.1 hit 38% in Aug 2025)(antigma.ai)
6 points by ubermon 5mo ago|9 comments
47.
How to Achieve #1 on Terminal Bench(twitter.com)
1 points by ubermon 7mo ago|1 comments
48.
by ubermon 7mo ago
and Why We Can't Have Nice Things): A Story
49.
by ubermon 1y ago
very good articulation on the problem. My bet is that it is probably fine in the sense that the mastery itself is not going to be relevant. I want to view AI coding as invention of new coding tools, at least in the way you described. I (hop
50.
by ubermon 3y ago
Give Copilot a try, it completely shift coding experience in Rust. Especially this: > I always have to search libraries and how to do things. Once you pass the initial curve with crutch like Copilot, then you can be almost as productive
51.
by ubermon 3y ago
haha was thinking the same. The other day openAI's api hiccup gave me a small panic attack
52.
by ubermon 3y ago
Moat does not come from compile time but runtime. The company, the operation, and the accumulated data, the brand, the trust and the reputation.
53.
by ubermon 6y ago
you seem really uninformed about what is happening in this space. Just to name a few, incredible fast settlement with lending and yielding; Automatic Market Maker; Decentralized Exchange guaranteed with no wash trade, insolvency and max tra
54.
by ubermon 6y ago
Curious on what is the general view of Indian population, twitter is full of trolls and paid campaign. From what I gather in chinese forum, I don't think any one believes China was the aggressive side, their most aggression is on getti
55.
by ubermon 6y ago
I think two important factors in play here, one is like what others said about better trust on the legal system in US to be relatively more independent; second I guess is just US market is too lucrative and significant to just let go.
56.
by ubermon 6y ago
I remember using this technique to get US content on Netflix while sitting in Canada
57.
by ubermon 6y ago
I think not manager but peers. They are mostly over achievers and somewhat competitive.
58.
by ubermon 6y ago
haha very possible, but it is still quite shocking to me that my mom, who is almost 60, keep telling me how fun it is...
59.
by ubermon 6y ago
Fueled by a recent Indian-China border conflicts, a political decision.
60.
by ubermon 6y ago
Could be some over zealous employees inspired by recent Indian-China conflicts and TikTok ban and decided within their own org to do it without realizing how big of a news it would be.

More ›