Sathwik

[Sathwik @sathwikmandava1]

@theo please stop this nonsense and build "sloper"

[Sathwik @sathwikmandava1]

@theo @davis7 I usually like your sponsors, but Paper's pricing is genuinely awful for actual users.

They give away unlimited editors for free, a feature companies with budget wouldn't mind paying for (and costs paper actual money). Then they cap MCP at 100 calls a week, which is the one thing a solo dev or hobbyist wants. And these aren't even calls to their servers. It's my local agent talking to their desktop app running on my own machine which costs paper almost nothing.

100 calls is useless. An agent burns through that building a single screen. So the "free plan" is really a demo, and the only way to use the core agent workflow you're promoting is $20/month to unlock something that costs them basically nothing to provide.

Charging individuals for local access while companies get collaboration for free is completely backwards.

[Sathwik @sathwikmandava1]

@theo What about AWS? I know it's more "getting-your-hands-dirty" than railway, but with Opus and AWS CDK, it should be a breeze, shouldn't it?

Honestly curious...

[Sathwik @sathwikmandava1]

@AnkanXplorer Did you just answer your own dumb fcuking question?

[Sathwik @sathwikmandava1]

@TimJayas If they roughly end up costing the same, what's the point?

[Sathwik @sathwikmandava1]

@jullerino @theo @t3dotcodes Stuck in here for past 30 minutes. Apps not opening and scared to quit this as I might mess with some background migration work...

[Sathwik @sathwikmandava1]

Not sure if anyone feels the same, but I still find Sol to be better at research (not coding, just architecture design and some other random research or lookups)...

Generally whenever I send a prompt to Chatgpt to look up something for me, I also send the same to Opus. With Opus I still see sometimes it wrongly states something with confidence, whereas Sol fact checks and catches itself almost always.

Don't get me wrong Opus 5.5 xhigh is still my main model, Sol 6.1 xhigh is my reviewer for coding tasks, but for general chat I find myself preferring/trusting Sol as I trust it's output more.

[Sathwik @sathwikmandava1]

@matthias_mrc If you are a Dev because you can write code, your plan A itself was stupid

[Sathwik @sathwikmandava1]

@Denche2005 Idiots

[Sathwik @sathwikmandava1]

My frustration is about lack of any open models from US not competing at any level with Chinese ones despite OpenAI and Antropic claiming to serve the needs of humanity not greed...

I agree frontier level intelligence without guardrails can be risky, but they should at least come up with (and maintain) some open weight models at least near to level of Luna or Haiku...

[Sathwik @sathwikmandava1]

Doesn't make sense. Competition (especially from open weights) forces the companies to train and server bigger/better models thereby purchasing more compute.

Nvidia LLMs exist for a different purpose than for general purpose assistants. Rather my frustration is about lack of any open models from US not competing at level with Chinese ones despite OpenAI and Antropic claiming to serve the needs of humanity not greed...

I agree frontier level intelligence without guardrails can be risky, but they should at least come up with (and maintain) some open weight models at least near to level of Luna or Haiku.

[Sathwik @sathwikmandava1]

@ArtificialAnlys Doesn't even compete in token efficiency with a much smaller 27b qwen model with almost twice the score

[Sathwik @sathwikmandava1]

@ArtificialAnlys Frontier is competing in 50s, China's open competing in 40s. US's open competing not even in 30s, but 20s...

[Sathwik @sathwikmandava1]

@JustinGorya @theo 80% capability at 10x reduction in cost might be a fair trade...

[Sathwik @sathwikmandava1]

@JustinGorya @theo People paying API prices might still be appreciating Sol's efficiency...

[Sathwik @sathwikmandava1]

@SamVale29 Thank you for asking me to be honest. I was not planning on being one, but seeing you mentioned to be honest I had a change of heart! 😇

Stupid fcuking "be honest" trend...

[Sathwik @sathwikmandava1]

Which rock were you sleeping under the past year?

Claude and Codex offer massive subsidisations (10x - 40x) on your subscriptions.

Cursor buys token usage from them at API prices (albeit a little subsidised) from those labs, which means you only get 1x - 1.5x usage of the money you spent.

[Sathwik @sathwikmandava1]

@ArtificialAnlys @cognition Sol 6.1 got overshadow by release of Opus 5.5, but any organisation paying API prices should seriously consider it as default model for most tasks...