I also have this problem!

It's the only model where an explicit instruction at the end of my message is sometimes ignored. This doesn't happen with any of the gpts, kimis, glms, qwen, etc. Just a deepseek problem.

Hope it improves!

I'm glad I'm not going insane...

I have also noticed this with Sonnet, funnily enough - it's not as strong, but it's still there. But yeah, I haven't seen this with any other model so far (although I mostly use the stronger ones - maybe it's a function of intelligence?).

Have you tried DeepSeek V4 Flash? It's very competent and extremely cheap.

I think Gemma 4 is also a good example of a capable small model.

I mention these not only because they're cheap but because they can run on consumer devices. The "every year bigger and more capable SOTA model" trend is mirrored by "the every year smaller and more capable open source model" trend.

256GB is what deepseek v4 flash with Q4 requires I believe. It is really still very far from “running locally on your device”. And it’s getting further away every day, looking at how the electronic market prices are surging.

I need to find stats on average RAM of personal devices, but I expect it will be so low, we are light years away from running a frontier model (from today) locally on a smartphone, let’s stop dreaming (and I really would love having it).

I do agree local models are progressing and I am to this day in awe at what a 50GB file can do – it still feels like black magic to me.

Also granted, something like Gemma 2 2B seems to have similar performance to ChatGP 3.5 and only require 2GB of RAM. But I think the RAM/performance ratio curve over time is logarithmic and not linear, it’s moving slower and slower.

Are you using Mimo 2.5 pro?

Yes. I tried a couple of weeks with non-Pro, and it was pretty good, but I had too many spare tokens, so I switched back to Pro. :)

I use it through my opencode go subscription and it's exactly how you described. Very pragmatic and not too ambitious. It's similar to Kimi 2.5/6 in that regard.

I like it!

Writing tests is also something AI agents excel at. At least they excel at converting plain english instructions into exact tests.

I haven't hand written tests in a while and it was something that I always bemoaned. Not anymore!

Azure suspended your account as well?

I think they meant that they migrated off of railway TO azure as opposed to FROM azure

But...AWS is a platform too, no? Seems like you're in the same category of risk you just moved to a more well-known name. Granted, Amazon is the most reliable even if they have their own quirks.

Each critical dependency you stack multiplies your risk. Now you have to worry about Railway AND Google causing business-damaging outages.

I was looking at this from Railway’s perspective. I really wonder what caused their account to be flagged, and they hint at more accounts being erroneously flagged as well.

Drilling in the basement seems like a pain to remove the dirt you dig up. Saving yourself a couple of feet cannot be worth the access troubles

This is spam - btw this is the first spam I have ever come across on hacker news

I think this was likely an attempted response to https://news.ycombinator.com/item?id=48008326

Yes - that’s got to be it.

FWIW, if you turn on "showdead", there is a ton of spam on HN. The mods are just really good.

Showdead is quite a disheartening experience - there’s just so much LLM generated crap. The dead internet theory doesn’t feel as fringe as it once did.

Oops I mixed up my tabs. My bad

And backups. Sqlite makes it easier but no backup process is easy. You always have to backup and restore at least once to have the confidence to rely on it.

It's another (big) point towards paying someone else to host it.

Its less of a worry given ts distributed.

Do you like Netdata? I'm looking into it. I'm curious if you use all the features or just a few.