VHRanger

8 ms·

91.
by VHRanger 1y ago
I'd also encourage you to try the !ai bang in addition to the "?" for quick answers. !ai sends the query directly to kagi.com/assistant with you default model. If you pick a solid and fast default, like kimi-k2, it'
92.
by VHRanger 1y ago
You are aware that if you're paying for kagi you can just slap !ai at the end of any search query and it'll direct the query to kagi.com/assistant, right? And that you can set the model to answer those query as any of the gro
93.
by VHRanger 1y ago
It uses only google as a backend, and the answer engine is based on the ancient llama 3.3 70B. It's well behind Kagi (which uses a bunch of sources for results and new models for quick answer or assistant). Perplexity is even well behi
94.
by VHRanger 1y ago
With ublock origin enabled on an android device I'd argue it's rather much faster
95.
by VHRanger 1y ago
You can trigger quick answer with a "?" But also feel free to try adding "!ai" to send the query to the assistant for a deeper dive
96.
by VHRanger 1y ago
Use kagi.com/assistant then? The cars still use that faster horse as a core engine before answering. Having a better engine makes for better answers keeping the model constant
97.
by VHRanger 1y ago
Or !ai to route it to kagi.com/assistant with your default model/agent to respond with kagi search results
98.
by VHRanger 1y ago
Kagi results consume brave search among others before returning the result so should be a superset in quality
99.
by VHRanger 1y ago
You can also use !ai at the end to redirect the query. It'll redirect the query to kagi.com/assistant and use your default model or assistant for the search. If you're paying for kagi, you can use ChatGPT or gpt 4.1 mini -- O
100.
by VHRanger 1y ago
You can also use !ai at the end to redirect the query. It'll redirect the query to kagi.com/assistant and use your default model or assistant for the search.
101.
by VHRanger 1y ago
They're far from SOTA: https://help.kagi.com/kagi/ai/llm-benchmark.html nova pro is worse than llama3-70B
102.
by VHRanger 1y ago
I bought an asus OLED zenbook 14 with the ryzen chip, slapped pop OS onto it and it ran with no issues since. In a lot of ways it's better than the M2 max macbook pro I had before (better screen for one). It was also, uh, 1/6th th
103.
by VHRanger 1y ago
Wouldnt something like this tool [0] be a cleaner solution? [0]: https://github.com/FlyGoat/RyzenAdj
104.
by VHRanger 1y ago
It's similar "parrot" behavior the models have on other inputs, even text. For instance, take the answers models have to this variation on the "surgeon son riddle": > A nurse comes to a surgeon and asks: "Si
105.
by VHRanger 1y ago
We're planning on open sourcing it soon anyways. We couldn't do it historically just because it was a dynamically generated string. But we're not, uh, hiding it - it's been posted on discord before as well.
106.
by VHRanger 1y ago
Do you have an example or two of a query that opus does well that others fail at?
107.
by VHRanger 1y ago
The difference is that Perplexity is subsidized by $0.5B in VC investment and are losing money. Kagi can only grow at the pace of incoming revenue.
108.
by VHRanger 1y ago
That policy is there because some users were incurring >$500 in AI costs in ultimate. Out of tens of thousands of users, the top 100 were using 48% of total AI costs. Kagi is bootstrapped, it can't just hemorrhage money like this th
109.
by VHRanger 1y ago
https://help.kagi.com/kagi/ai/llm-benchmark.html gemini flash, 4o-mini, mistral small for super cheap qwen-qwq for strong models
110.
by VHRanger 1y ago
Should be out globally today actually
111.
by VHRanger 2y ago
Oh wow, those authors. For those not in the know, Carl Shapiro [1] and Hal Varian [2] are well known economics researchers. Hal Varian ended up working at Google building the AdWords auction after writing this it seems. For anyone wondering
112.
by VHRanger 2y ago
VPNs would not make sense if the actual ISP if blocking you. Apart from something like starlink and even then they're not playing nice with geographical access most of the time based on politics, whims of a narcissist, or just business
113.
by VHRanger 2y ago
I mean, if there's restricted supply/excess demand then the ones at the margin moving the point where supply and demand cross will the ones willing to pay high prices. It shouldn't surprise anyone who took econ 101 that new b
114.
by VHRanger 2y ago
I draw the line at 5 personally. 4 if I'm being picky. A mere 1,000x difference? This is just pedantry
115.
by VHRanger 2y ago
what happens when you prioritize loyalty over all else. This hiring and promotion pattern is also why autocracies tend to be mismanaged nightmares
116.
by VHRanger 2y ago
Right, so you're grading it on bean counting. Importantly, the NASDAQ also had a 9-10x increase in this period, so while Apple did grow, it's not exploding benchmarks under his tenure as far as I see it.
117.
by VHRanger 2y ago
Tim Cook is a bean counter. His one significant win as a CEO was making the profit margins on apple processors higher.
118.
by VHRanger 2y ago
While this is a neat sounding idea, there's a few issues here: 1. Embeddings are very close to a reversible function. It's not hard to take an embedding and query back the closest query semantically from the source LLM. 2. We alre
119.
by VHRanger 2y ago
I would include helix in the list with nvim nowadays. The LSP support is native and very easy to set up for most languages.
120.
by VHRanger 2y ago
Can you explain what this does? From screenshots I can't tell what the ddterm user experience is I personally tend to wezterm nowadays, kitty breaks too many things, especially over tmux/ssh (or combination thereof and with zellij

More ›