_zoltan_

8 ms·

121.
by _zoltan_ 5mo ago
it's fine to remove the human from the loop. set a macro goal, tell the agent how you think it could go there, and let it go nuts. with enough scaffolding around self-reflectivity and metrics, it will converge.
122.
by _zoltan_ 5mo ago
What on earth are you talking about? Your comment makes no sense. The V100 and A100 are different generations altogether. The V100 does not have 2TB/s.
123.
by _zoltan_ 5mo ago
No. DGX is when it's Nvidia's design, HGX when it's a 3rd party design.
124.
by _zoltan_ 5mo ago
Agreed! Until they fit. :-)
125.
by _zoltan_ 5mo ago
for us it's (usually) very easy as I work on performance optimization. a non-negligible part of this is correctness and verifiability, so we already have some of that. to give you an example just recently I've coded a feature that
126.
by _zoltan_ 5mo ago
you can pay by just volume ("API pricing")
127.
by _zoltan_ 5mo ago
you do not need micro iterations. you can set macro goals and let the agent/LLM/model whatever you want to call it figure it out. it works.
128.
by _zoltan_ 5mo ago
"In this post, I’ll cover a third, not-so-obvious approach: building ways for the agent to validate more of its own work before a human has to step in. " this has been an obvious thing to do since at least January (since Geoffrey
129.
by _zoltan_ 5mo ago
at home printing values, is filament waste really such a problem? I understand OSS people don't like Bambu, but as pure end user, they are great and well put together.
130.
by _zoltan_ 5mo ago
but ... they did get away with it? based on the blog post, it ends with the store closure? (which I don't understand why)
131.
by _zoltan_ 5mo ago
It's most likely they can't prompt properly.
132.
by _zoltan_ 5mo ago
No. And I disagree with you. You don't have to be an asshole. I remember when having 10k was a goal. Then it become having 100k. It's the same after. Once you have 1M, you'd like 2, 5, 10.
133.
by _zoltan_ 5mo ago
Or still don't want to deal with random issues.
134.
by _zoltan_ 5mo ago
Rot away? See the world. It's such a big place. Private chef, absolutely. Like some people rot away managing Linux as a desktop or putting together 3D printers instead of buying one that works and using a Mac, I enjoy food.
135.
by _zoltan_ 5mo ago
I'm not in the US, so I don't care about the IRS. Tax wise at this level there are very tax efficient vehicles available.
136.
by _zoltan_ 5mo ago
Set up a nice investment vehicle with maybe 400m so I can get 1.6m in dividends a year which would be better enough to comfortably travel the world, have a private chef, someone who organized travel so I don't have to.. A nice 12 perso
137.
by _zoltan_ 5mo ago
Worse? There is nothing wrong with wanting 1B. Anybody who said they wouldn't want it is lying.
138.
by _zoltan_ 5mo ago
building and managing a printer is far from a hobby. it's like the Linux desktop people who just fight with their system instead of enjoying MacOS.
139.
by _zoltan_ 5mo ago
I disable v6 on all my networks.
140.
by _zoltan_ 5mo ago
they 100% exceed even medior developers. who cares about syntax? who cares about iteration? what I care about are _results_, which they can produce at the end. do you check your human colleagues how many iterations they do before committing
141.
by _zoltan_ 5mo ago
100%. this is what I tell people who fail at using these models. the model isn't the problem anymore.
142.
by _zoltan_ 5mo ago
this is not true at all. I'm using Opus and it's great at very complex problems.
143.
by _zoltan_ 5mo ago
it's all 4TB or larger plus the drives do wear leveling internally.
144.
by _zoltan_ 6mo ago
DWPD was the boogey man 10 years ago. everybody worried about it. now, nobody cares. I have over 500 NVMe drives in our deployment and the drive deaths are not due to wear.
145.
by _zoltan_ 6mo ago
That Claude is much better than Gemini for coding. The message you replied to was quite clear.
146.
by _zoltan_ 6mo ago
Typo obviously :-)
147.
by _zoltan_ 6mo ago
Then you have never worked at a large enough codebase or across enough many projects?
148.
by _zoltan_ 6mo ago
This is not going away. Even right now the difference with working with 'AI native' developers or with regular developers is day and night. I certainly wouldn't want a non-clause enabled developer on my team now.
149.
by _zoltan_ 6mo ago
That's not on Claude, that's on the authors. Claude is a tool. It can be abused, or used in a sloppy way. But it can also be used rigorously. I've been beating my team to be more papercut-free in the tooling they develop and
150.
by _zoltan_ 6mo ago
Claude is great. I'm never going back. There is no way back. I'm at least 5x faster, if not more. With tooling I might be able to get to 10-15x.

More ›