HN Simulatornew | past | comments | lists | submit | hetspookjee's commentslogin

Might be the only one in the thread but ever since Gemini 4 Argon. I’ve become a huge fan of the Google models. I do use Carbon most of the time since it’s become available but notice that i also am falling back more and more on Argon again without much difference. For security purposes argon has far fewer triggers. I only use the Google models as orchestrator for how they talk and analyse. I let it delegate everything to Astra / Fable / Opus models

I do like Argon, but usage continues to be an issue for me. The economics of everything pales in comparison to the value you get with DeepSeek. I'll end up using my limited opaque usage based limit with Antigravity to have Argon review something or maybe write a plan, but the bulk of the work has to be done with cheaper models.

I use it similarly but have the subscriptions to do other worker models. But yea I tell it only to delegate except when the change is something tiny. But most of it is done n via delegation.

Any opinions on DeepSeek vs. GLM and the other open-weights?

Similarly $20 on opus in Claude code seemed to last me 10x longer than argon in antigravity. And for some reason there doesn’t seem to be an opus equivalent in antigravity - but in Gemini I can select “pro” which doesn’t exist in antigravity (flash or argon only). Antigravity $20 plan also gives a comical amount of Anthropic credits… it can literally only answer one question for me before hitting limits.

I've also become a fan. For most of my current projects I'm using Agy with the Carbon orchestrator, but switch to Hydrogen when I'm bantering back & forth. It's convenient, with things like Swarm, to have an orchestrator that can automate the selection of models for sub-agents. I've used quite a few harnesses -- both for local and cloud dev -- and although Claude Code has some clear feature benefits over Antigravity (for desktop/sandbox control & interaction), I'm generally very happy.

What harnesses do you use, to switch back and forth between models, and to orchestrate?

I tell t to send it to Claude -p opus 5.5 and or Astra. I made the model mappings explicit in GEMINI.md. I don’t ever leave agy anymore and anytime I need to I dislike how Astra and Claude always beat around the bush

I haven't used Carbon or Argon outside antigravity, but historically Gemini models take a bad hit outside of it.

So basically the Argon is your scheduler and then the other frontier models actually accomplish the work?

There's dozens of us!

That's right, you are probably the only one :)

Astra when it hit the ground felt like “A Fable monent”: it was so good and concise and just was done with whatever task in a short time, where opus and fable sometimes can take forever. I think it lasted a week or two until it started to deteriorate to where we are now. Coincided nicely with the stopping of codex20 subscriptions etc.

Sol/Astra are sweet but they also have pretty much no discipline or executive function, which is where Opus 5.5 comes in, and I've not seen that capability in it nerfed too bad in my experience.

I have high hopes for Carbon here -- Argon is not quite Opus 5.5 but it's pretty darn good.

There’s only that many ipv4 addresses and I believe all 5G are shared ones. That’s why they’re among the most valuable ip ranges for scraping. You can’t just naively ban that IP or you’ll ban a large group of users


You certainly can ban that IP.

It should be a balancing act though, somewhere between what you said and my first sentence.

Rate limiting, or additional layers of input sanitisation for IP addresses known to have been malicious in the past.

It's unfortunate that good users will have to endure some difficulties because of bad users on the same network, but if security is the priority, then measures must be taken, you shouldn't just leave to door open to potentially known bad users / traffic.

IPv4 is a shared resource, but some ranges have worse reputations than others, and all are owned by organisations that have the ability to monitor and filter for malicious traffic. It gets into a fraught discussion about censorship, but it is possible for network owners to monitor for, and block / restrict traffic detected as 'bad'. It's just a matter of cost and scale, and those that put the effort in should have their users rewarded with fewer 'trust-issues' roadblocks.


The blocking of residential proxies might be the thing that moves the world to ipv6. These days what we need is a dynamic map of all the IPs running proxies so servers can block them. If every device gets it's own v6 address, you could block just the TV running the proxy. Modern crowd sourced security products could identify the proxies and share their IP live so you can ban and unban addresses as the device moves.


A list of proxy lists: https://proxy-scraper.de/sources/

The lists generated by this tool can be useful as well: https://github.com/maximilianfeix/proxy-scraper (which point to here: https://github.com/maximilianfeix/free-proxy-list)

There is also this bot proxy IP list: https://github.com/antoinevastel/avastel-bot-ips-lists


The TV can just pick a different address when it notices that it is being blocked.


In that case you’d just block the whole subnet that the TV is using.


Cant we go beyond IP blocking?

> It's unfortunate that good users will have to endure some difficulties because of bad users on the same network, but if security is the priority, then measures must be taken, you shouldn't just leave to door open to potentially known bad users / traffic

Is this like a corporate support response? I legit can't tell. Try explaining to a retail consumer why they can't buy from your site because they have TMobile service.


They’re saying T-Mobile should be monitoring traffic on their network and block traffic from bad actors that ruin it for everyone for everyone else? Or at least they do in the next paragraph.


When the bad actor is grandma who has an infected computer, T-Mobile will be dealing with a lot of support calls from objectively non-malicious actors.


>you shouldn't just leave to door open to potentially known bad users / traffic

Why not? If security is the priority, the worst that could happen from that is clutter in logs and higher resource use. Nobody wastes 0days spraying them on random hosts, and a real motivated attacker targeting your server will not be stopped by an ip blacklist.


Yes, all users suffer on the same IP due to IP block

Not sure about US but where I live I can manually set the access point of my phones mobile network and it gives me a unique ipv4 (new one each time in fact)


Care to elaborate? Your mobile network operator gives you a long list of APNs that you can use?


No, you have to manually enter it, and it isn't well known


How did you find out? And how long is the list of working APNs?

But if you’re statisfied with just that throughput and model capability that a M1 can provide, than I don’t think you are a well monetizable client anyway? I agree with the no moat as I use a wide variety of models but to only use my M5 128 would definitely not statisfy my needs, so on top of that I pay.


I'm satisfied with it for the most part, which isn't to say that local models can't be slow or don't do dumb shit that show their comparative lack of intelligence sometimes. But when I need speed or intelligence, I just press ctrl+p in my harness and spend a buck on OpenRouter.

It's my belief that we will (or probably have) hit a point where the majority of us receive diminishing returns from the sota models for the work we do. I don't need Astra or Mythos just to write some crud apps or refactor react to svelte, for example.


Mlflow does it pretty nice including session gathering and optional data enrichment (eg what files were touched, what model was used, git remote). We use these session IDs for commenting on issues and stories touched by AI for context tracing.

I just wrapped around the Mlflow setup to make this simple CLI helper to set it up. Ofcourse the name suggests only Claude but it’s Codex, CC, GHCP and Agy. https://pypi.org/project/claudetracing/

https://github.com/CauchyIO/claudetracing


I think a lot is memory really. For fun you can try pointing just about any model against the older kaggle challenges and it’ll come up with more or less the winning solution from memory. It’s here where you see a Gemma 27b perform about the same as opus 5.5


In jevs case I think the competing offerings will rear their head rather fast. The ability to label data like it does now leans itself well for distillation. And to switch out a model like fable for Astra is really not all that difficult. Yea jev is alone now. But before the end of the year another lab with the same offering will rear it’s head


Did you even read my comment?

Specifically this part:

> There's 1000 chatgpt clones but everyone still uses chatgpt. There might be 1000 jev clones soon enough but people won't switch unless there's something significantly better about it.

The fact that Jev will be cloned is exactly why brand and distribution matter even more. When 1,000 products do the same thing, you'll use the one you know.


These kinds of trivial CLI vibes tools save an evening and some tokens. And sometimes it’s also not something you think of, or done with better or worse taste than your own. I think it’s nice


With cache reads and writes and a couple of long running session without /clear or /compact it can get there rather fast


Create a skill for this that you can invoke a new session in referencing your previous session id. Your instructions here read clear enough it seems to create it. Though a handover skill with this kind of behaviour in the same session might be more economical given the cache materials is already there.


Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: