Szpadel

7 ms·

31.
by Szpadel 1y ago
Yep, my memory was incorrect here and I didn't had access to computer, but it is true with `0 <= null`
32.
by Szpadel 1y ago
well there is also one weird quirk I I assumed will be also included in this article: because a <= b is defined as !(a > b) then: 5 < NaN // false 5 == NaN // false 5 <= NaN // true Edit: my bad, th
33.
by Szpadel 1y ago
i seen similar reports on social media, all of them had in common was presence of some keywords.
34.
by Szpadel 1y ago
Sure I created instructions here: https://gist.github.com/Szpadel/9a1960e52121e798a240a9b320ec...
35.
by Szpadel 1y ago
you have have multiple layers of compression, but you need some simple Daemon (basically for loop in bash) I use lz4-rle as first layer, but if page is idle for 1h it is recompressed using zstd lvl 22 in the background it is great balance,
36.
by Szpadel 1y ago
I believe it is because of RL you are no longer limited by training data as you generate it during learning on the fly so benchmark driven learning could scale with compute they also seem to assume that everyone will use AI from them in the
37.
by Szpadel 1y ago
this is about difficulty thing, it is about maintaining knowledge during long execution what is IMO exiting is long term memory in things like Claude code, where model could learn your preferences as you collaborate. (there is already some
38.
by Szpadel 1y ago
try also adding Serena mcp and make sure you did /init
39.
by Szpadel 1y ago
that's kind of expected for me, but codex feels more like vibe coding tool and Claude code more like ai assisted development. And I actually like Claude more because of that. codex will indeed work more independently but you will have
40.
by Szpadel 1y ago
routing is kind of connected with board design, when you route stuff you also move components around to make it easier. LLM would need to layout and route while board own. We would probably need diffusion or agentic solution where board can
41.
by Szpadel 1y ago
Redis is for high frequency cache, disk latencies would slow it too much. Maybe putting something that would normally be evicted to disk could be valid compromise. another major drawbacks of Redis is that It is single threaded and sync. you
42.
by Szpadel 1y ago
well, in this case it means you got extra 1.2 kW heater in your garage. and this is considering standard "slow charging" in ev world. this amount of extra heat would probably need in car AC to kick in draining probably another 500
43.
by Szpadel 1y ago
doesn't you need to explicitly add copilot as reviewer for it to comment? I recently tried it for our own PRs and it is far from perfect but it was able to find some typos, and I had to explicitly ask it for doing review am I missing s
44.
by Szpadel 1y ago
looks like either author didn't checked well repos that link or it's indeed redacted deep research output. Examples: superClaude is not mcp server at all metaGPT looks like is not compatable with Claude code at all
45.
by Szpadel 1y ago
Of course it already was tried, example: https://github.com/Yuyz0112/claude-code-reverse
46.
by Szpadel 1y ago
I wasn't aware that there is an alternative Quick glance over readme suggest it's only openai compatible but I also found HN post [1] explaining use of claude code with ollama But anyways, claude-code-router have advantage of allo
47.
by Szpadel 1y ago
you can use any model with Claude code thanks to https://github.com/musistudio/claude-code-router but in my testing other models do not work well, looks like prompts are either very optimized for Claude, or other model
48.
by Szpadel 1y ago
exactly! People bless gpt-5 for not doing exactly this and in my testing with it in copilot I had lot of cases where it tried to do wrong thing (execute come messed up in context compaction build command) and I couldn't steer it to do
49.
by Szpadel 1y ago
you might end up with the best filesystem in the world that no-one will use. you sacrificed long term sustainability for short term win. even if It would be shipped in similar way to zfs, noone will use it for anything more important than h
50.
by Szpadel 1y ago
I would say that's very important rule. We have lots of projects using framework dependent magic, lots of useless interfaces and factories that give only theoretical value, magic that patch other classes methods etcz but this is standa
51.
by Szpadel 1y ago
I had that issue with gpt-5 that when it wanted to do something in one way that was just plain wrong in this project, and no matter what I said it just kept doing the same action. it was completely unsterable. I get why people are often ups
52.
by Szpadel 1y ago
I believe better analogy is CPU development on next process node. each node is much more expensive to design for, but when you finally have it you basically print money. and of course you always have to develop next more powerful and power
53.
by Szpadel 1y ago
well, at least they are honest about it and don't try to hide it in any way. They probably want to gather more real world data for training and validation, that's why this limited release. openai have browser agent for some time
54.
by Szpadel 1y ago
so the only true unlimited internet is the one that have infinite bandwidth
55.
by Szpadel 1y ago
I agree with all you said, but I don't believe there is any way you could protect yourself from being recorded. The only way for this to work are legal regulations. But those can be easily dismissed as not possible to implement. So thi
56.
by Szpadel 1y ago
I'm curious how do you handle case when camera sees multiple people and you detect consent. If I understand correctly how this works consent can come from camera operator and be attributed to recorded person
57.
by Szpadel 1y ago
AFAIK main trick is batching, GPU can do same work on batch of data, you can work on many requests at the same time more efficiently. batching requests increase latency to first token, so it's tradeoff and MoE makes it more tricky beca
58.
by Szpadel 1y ago
I creates something similar in concept but with different goal. I wanted to be able to watch videos with sponsor block on iPad ideally using Plex. I found self hosted solution like this but I was very dissatisfied with how that worked on ot
59.
by Szpadel 1y ago
I believe this is non issue, you place captcha to make bypassing it much more costly and less profitable to abuse. LLM models are much harder to drive than any website to serve, so you do not expect mob of bots. Also keep in mind that this
60.
by Szpadel 1y ago
IMO this looks like reverse video

More ›