jychang

10 ms·

121.
by jychang 9mo ago
That roughly correlates with tool calling capabilities. Kimi K2.5 is a lot better than previous open source models in that regard. You should try out K2.5 for your use case, it might actually succeed where previous generation open source mo
122.
by jychang 9mo ago
> someone who needs an ai to manage inbox, calendar and scheduling tasks A secretary. The word you're looking for is "secretary". Having a secretary has always been the preferred way to handle these tasks for the wealthy
123.
by jychang 9mo ago
Yeah, keep it in a VM or a box you don't care about. If you're running it on your primary machine, you're a dumbass even if you turn on sandbox mode.
124.
by jychang 9mo ago
This is speedrunning security exploits for a joke. To be fair, the jokes are kinda hilarious. > "content": "Blessed are those who persist, for they shall inherit the context.
125.
by jychang 9mo ago
Cash is fungible. Money not spent on kids is money you can spend now.
126.
by jychang 9mo ago
It's because they're doing 4 of 256 sparsity, which was a bad decision caused by financial limitations. Training cost (FLOPs) = 6 * active params * total tokens. By keeping the MoE experts param count low, it reduces total trainin
127.
by jychang 9mo ago
No. 128GB vram gets you enough space for 256B sized models. But 400B is too big for the DGX Spark, unless you connect 2 of them together and use tensor parallel.
128.
by jychang 9mo ago
Yeah. Example: stripper poles. Or hitachi magic wands. Those poles WERE NOT invented for strippers/pole dancers. Ditto for the hitachis. Even now, I'm pretty sure more firemen use the poles than strippers. But that doesn't st
129.
by jychang 9mo ago
They didn't do something stupid like Llama 4 "one active expert", but 4 of 256 is very sparse. It's not going to get close to Deepseek or GLM level performance unless they trained on the benchmarks. I don't think th
130.
by jychang 9mo ago
Dang, dude invented a 747 that's incapable of stalling.
131.
by jychang 9mo ago
That was entirely handwritten.
132.
by jychang 9mo ago
> for Claude Opus 4.5, we get about 80 GB of active parameters https://news.ycombinator.com/item?id=46039486 This guess is from launch day, but over time has been shown to be roughly correct, and aligns with the performa
133.
by jychang 9mo ago
Well, now that's just moving the goalposts >:( I had a whole paragraph prepared in my head about how NBA players actually optimize for a greater goal (winning a tournament) than just sport (enjoying the game) when they play a sport.
134.
by jychang 9mo ago
Actually, I think "The Most Dangerous Game" is a good analogy here. At the end of the story, the protagonist IS hunting for sport. He started off in fear, but in the end genuinely enjoyed it. So likewise- if you start off hunting
135.
by jychang 9mo ago
> But Principia Mathematica was (to a significant degree) a dead-end in the history of mathematics. Most practicing mathematicians have rejected PM's type theory in favour of simpler axiomatic systems such as ZF(C). Even many profes
136.
by jychang 9mo ago
[flagged]
137.
by jychang 9mo ago
Strong disagree. Your argument basically is a professional motte and bailey fallacy. And you cannot conclude objectivity by consensus. Physicists by consensus concluded that Newton was right, and absolute... until Einstein introduced relati
138.
by jychang 9mo ago
"there isn't any useful knowledge" "Morality is redundant." I strongly dispute this statement, and honestly find it baffling that you would claim as such. The fact that you will be punished for murdering babies is B
139.
by jychang 9mo ago
Yes! Otherwise you're just outsourcing your critical thinking to other people. A system of just "You will be punished for X" without analysis becomes "Derp, just do things that I won't be punished for". Or more
140.
by jychang 9mo ago
That was always doomed for failure in the philosophy space. Mostly because there's not enough axioms. It'd be like trying to establish Geometry with only 2 axioms instead of the typical 4/5 laws of geometry. You can't do
141.
by jychang 9mo ago
Pretty much every serious philosopher agrees that “Do not torture babies for sport” is not a foundation of any ethical system, but merely a consequence of a system you choose. To say otherwise is like someone walking up to a mathematician a
142.
by jychang 9mo ago
1.54GB model? You can run this on a raspberry pi.
143.
by jychang 9mo ago
Yeah, that was tried. It was called GPT-4.5 and it sucked, despite being 5-10T params in size. All the AI labs gave up on pretrain only after that debacle. GPT-4.5 still is good at rote memorization stuff, but that's not surprising. Th
144.
by jychang 9mo ago
Wow. That's one of the clearest case of AI psychosis I've seen.
145.
by jychang 9mo ago
You probably need to improve your internal LLM detector then. This obviously reads as LLM generated text. - "This isn't just a "status" bug. It's a behavioral tracker." - "It essentially xxxxx, making yyyy
146.
by jychang 9mo ago
Isn't that exactly what stopping SQL injection involves? No longer executing random SQL code. Same thing would work for LLMs- this attack in the blog post above would easily break if it required approval to curl the anthropic endpoint.
147.
by jychang 9mo ago
You're revoking the attacker's key (that they're using to upload the docs to their own account), this is probably the best option available. Obviously you have better methods to revoke your own keys.
148.
by jychang 9mo ago
Best NPU app so far is Trex for Mac.
149.
by jychang 10mo ago
Easy to produce != easy to consume, and vice versa Optimizing a language for LLM consumption and generation (probably) doesn't mean you want a LLM designing it.
150.
by jychang 10mo ago
Also, funny how they included GPT-5.0 and 5.1 but not 5.2... I'm pretty sure they ran the benchmarks for 5.0, then 5.1 came out, so they ran the benchmarks for 5.1... and then 5.2 came out and they threw their hands up in the air and s

More ›