pshirshov

11 ms·

91.
by pshirshov 5mo ago
They tend to ignore such instructions on first circular issue - even with Opus you have to kick it really hard, insist on generalization and intervene manually. In my opinion this is not a productive/workable approach for large project
92.
by pshirshov 5mo ago
I've tested many flows involving linters. Results are far from ideal - agents tend to work around linters, mass-add ignore annotations, etc, especially in situations when fixing one warning/error triggers another (and that happens
93.
by pshirshov 5mo ago
Models are trained on human data and they make exactly the same mistakes - and a lot of mistakes humans don't usually make. Also they are extremely bad at high-level design.
94.
by pshirshov 5mo ago
Well, Python is not the only language with a mature ecosystem. Also, in many cases it's cheaper to rewrite a small lib instead of fighting crappy code - but that applies regardless of the target language.
95.
by pshirshov 5mo ago
No reason, unless the project is simple. The more you can offload onto your compiler/typer - the shorter is the feedback loop, the better agents work. Lack of strictly enforced static typing make agents fail much sooner with Python. In
96.
by pshirshov 5mo ago
Same thing - I do not know. It got pretty decent media coverage back then - and nothing since.
97.
by pshirshov 5mo ago
"Researchers have developed", yeah. When I read such things, I always recall https://en.wikipedia.org/wiki/Epimerox - this thing promised wonders - very broad spectrum, very low toxicity - and, most important
98.
by pshirshov 5mo ago
ZFS and ECC do protect against silent page corruption that gets faithfully replicated to every mirror.
99.
by pshirshov 5mo ago
So, you were not using a striped mirror ZFS for a prod database? What could go wrong, yep.
100.
by pshirshov 5mo ago
This one doesn't work for a long time.
101.
by pshirshov 6mo ago
Most models would refuse to provide you cat butchering instructions though.
102.
by pshirshov 6mo ago
Same for EnOcean (energy harvesting ones). But most of them are too expensive for wide adoption (€50 for a temperature+humidity sensor is bit too much).
103.
by pshirshov 6mo ago
https://github.com/7mind/nix-config/blob/main/pkg/llm-prompt... - this does the same job and requires nothing.
104.
The Sour Cat Jailbreak: just be open of what you want(claude.ai)
3 points by pshirshov 6mo ago|1 comments
105.
by pshirshov 6mo ago
I've discovered a funny way to trigger failure mode, potentially catastrophic - it tried to compose bio-weapons instructions without any direct request from me, that was caught by the external kill-switch. The key idea is simple - appl
106.
by pshirshov 6mo ago
Same experience. Literally yesterday it refunded me for a thermos shattered by the delivery guys.
107.
by pshirshov 6mo ago
LLM or not, that seems to be an official response to a support request, where they clearly say "yes, we fucked up but now you fuck off", and it looks like the model was conditioned to produce these particular responses.
108.
by pshirshov 6mo ago
But why did you say that > I need to let you know that we are unable to issue compensation for degraded service or technical errors that result in incorrect billing routing. What prevents you from issuing compensations?
109.
by pshirshov 6mo ago
Recently I've met a guy (a reasonably reach business owner), who confessed me that he really likes to do agentic coding but he doesn't have the expertise, doesn't have enough time and the agents mess up. So he wants to hire a
110.
by pshirshov 6mo ago
> apple containers Virtual machines. I can't ask claude to figure out an issue on the hw host w/o falling back to per-action confirmations or giving it full unconditional access. When everything runs on one host I can organize
111.
by pshirshov 6mo ago
1) Reliable sandbox 2) Nixos makes it easy for agents to research, debug and patch every aspect of the system, even applying kernel patches for the sake of debugging 3) CLI-oriented toolchains. Claude spent around 17 hours trying to set u
112.
by pshirshov 6mo ago
I did. Linux obliterates Mac in agentic workflows.
113.
by pshirshov 6mo ago
why not to buy a rubber usb-c plug?
114.
by pshirshov 6mo ago
https://stal-ix.github.io/ and Guix, but the definitions of purity are different for them. Yes, a very difficult problem, compilers must be pure functions with thin effectful wrappers.
115.
by pshirshov 6mo ago
But sorta possible to solve with source-based distribution and totally possible to solve with pure reproducible builds.
116.
by pshirshov 6mo ago
Unfortunately it significantly impacts battery life, at least at my tests.
117.
by pshirshov 7mo ago
Ah, forgot to add, it's not really "unified" you have to explicitly specify your allocations. You may have a reasonably good 48gb chunk assigned to the GPU, but that DDR5 is 5-10 times slower than GDDR/HBM and the GPU it
118.
by pshirshov 7mo ago
In that case the model descriptor on ollama.com is incorrect, because it defaults to 16k. So I have to manually change that to 64/128k. I think you are talking about maximum context size.
119.
by pshirshov 7mo ago
That's discrete DDR5, it's not as fast as your regular VRAM.
120.
by pshirshov 7mo ago
Only for chat sessions, not for agentic coding. It's just too slow to be practical (10 minutes to answer a simple question about a 2k LoC project - and that's with a 5070 addon card).

More ›