try-working

4 ms·

31.
by try-working 1mo ago
Generally you should only have two models in the pool per domain. I wrote some of my learnings building a router here: https://try.works/first-principles-of-model-routing
32.
by try-working 1mo ago
I built a router that lets you route between local and cloud models. Link in my profile.
33.
by try-working 1mo ago
Laptops can't do agentic engineering. They get hot as hell and battery drains instantly. I think this will promote a switch to desktops for the next couple of years, until we have new mobile chips.
34.
by try-working 1mo ago
Twitter users.
35.
by try-working 1mo ago
Nvidia
36.
by try-working 1mo ago
Not sure I would trust Luna with that. Deepseek Pro Max and Code Mode I would be more inclined to trust.
37.
by try-working 2mo ago
You could let a agent drive the browser verification
38.
by try-working 2mo ago
Margin collapse
39.
by try-working 2mo ago
I main V4 Pro at work now, and at home I route between Pro and Flash based on task. Switched to Opus 4.6 for some tasks at work because I needed image input - horrible. So nice to get image input with DS. Edit: I see it has limited resoluti
40.
by try-working 2mo ago
So their only answer to DeepSeek V4, Kimi K3, GPT 5.6 price cuts on Luna (80%) and Sol (50% on OpenRouter) is to temporarily extend usage limits, then keep extending past the deadline every two weeks because they still don't have an an
41.
by try-working 2mo ago
that seems to be how most Chinese models achieve increased benchmark scores. GLM and Kimi models are "thinkslop" models that reason over their own thinking, which increases cost and decreases speed significantly. That's why G
42.
by try-working 2mo ago
I've been thinking that they should design the DeepSeek harness to work with other providers since a large use case is to off-load work from an expensive model to DeepSeek, instead of makign everyone hack the harness. I see that it wor
43.
by try-working 2mo ago
edit: updated the answer above to be more qualitative instead
44.
by try-working 2mo ago
To keep it simple, forget about routers and imagine you're in Cursor using GPT for a while, reaching a cache of says 200k. You decide to switch to DeepSeek in the same session via the model picker, and continue as usual. What happens i
45.
by try-working 2mo ago
No, you still don't understand what I'm saying. Yes, ASICs make inference faster, but also makes the hardware obsolete faster, if it's embedded in a phone. That sounds like a negative, but Apple is going to turn it into a pos
46.
by try-working 2mo ago
What I'm saying is that Apple will use these type of models etched into chips, and they will do it because it drives obsolescence, so they can shorten the upgrade cycle. They will do it because they figure out it's good for them.
47.
by try-working 2mo ago
What's even more noticable is that Anthropic still hasn't responded to the Kimi K3 release or the DeepSeek release.
48.
by try-working 2mo ago
you don't understand what I wrote.
49.
by try-working 2mo ago
obsolescence is the whole point. apple gets to sell a new phone very 6-12 months because of it. i have written about this: "For device makers Packaging models with laptops and smartphones will let application access near free, low late
50.
by try-working 2mo ago
Flash is the most used model in the world since last week
51.
by try-working 2mo ago
yes, and this is why we need model routing
52.
by try-working 2mo ago
it's 200 rmb and it looks super fun. perfect product.
53.
by try-working 2mo ago
I recommend the Shadowrun series on Steam. Not large games but really fun.
54.
by try-working 2mo ago
agree. what people dont understand is that the blacklisting is what created motivation for investing capital, manpower and years of time into these companies, or initiatives inside existing companies. most of this would have never gotten of
55.
by try-working 2mo ago
this is not a good benchmark for models, but it's great if you're optimizing for attention on twitter because video content and 3d animations perform best on social media. a real benchmark is instead running evals on your own trac
56.
by try-working 2mo ago
No, you are wrong. Without the blacklisting and export bans the DUV machine in the article would not exist today.
57.
by try-working 2mo ago
All of this can be traced back to the 2018 blacklisting of ZTE, and the continued willful ignorance of China's capabilities, and of the history of industrialization and technological progress.
58.
by try-working 2mo ago
that's why i recommend running everything through a router.
59.
by try-working 2mo ago
the repo is here and you can also find my twitter in my profile: https://github.com/try-works/role-model you can also read this: https://try.works/first-principles-of-model-routing cache hit rate is ke
60.
by try-working 2mo ago
yep. I've been saying for month that exactly this will happen with DeepSeek and Kimi K3, and margins will erode, and Anthropic will fail to IPO.

More ›