Ruthvik Dev

typescript SWE & founder.

[Ruthvik Dev @ruthvik_devx]

@ptrschmdtnlsn the opposite happened to me except my banking app showed my account balance as being a few hundred lower than what it actually was

thought for a moment that I had bought something by accident or got scammed

[Ruthvik Dev @ruthvik_devx]

@sidkasbekar @airtreevc sf looks great

also nice to see a photo of sf that isn't the typical photo standing in front of the golden gate bridge

[Ruthvik Dev @ruthvik_devx]

scrolling through your AI chat history gives a nice timeline of what you were doing and when, as well as your mindset and priorities at a given point in time

[Ruthvik Dev @ruthvik_devx]

@ChrisWillx @robkhenderson the US also does a great job of attracting millionaires to move there as well

it is a very effective positive feedback cycle

[Ruthvik Dev @ruthvik_devx]

@zdogmode too variable

engineering problems, and their respective solutions either work or don't, but marketing is much more dynamic and with no 1 correct solution

[Ruthvik Dev @ruthvik_devx]

@hlthisholistic @dhh was strongly considering getting a XPS as my next laptop, but leaning a bit more towards the framework after looking more into it

Dell's Enterprise range is also great from what I've heard

[Ruthvik Dev @ruthvik_devx]

macs are great for local AI, but unified memory still has tradeoffs compared with traditional VRAM on NVIDIA GPUs

NVIDIA is better for training tho because of raw compute + CUDA and specialised low-precision hardware

but for inference on very large models, especially MoEs with total weights > normal VRAM capacities, Macs are the value king

[Ruthvik Dev @ruthvik_devx]

@MaruPelkar I recall 0x Alpha using the term "load-bearing" as well as a few other Claudisms

just based on that, I don't think it is was originally developed by a US lab but there is a decent likelihood it is a fine-tune of GLM-5.3

[Ruthvik Dev @ruthvik_devx]

one thing that doesn't seem to be discussed as much with the new generation of "personal AI computers" is how different memory capacity and inference performance are becoming as constraints

NVIDIA's DGX Spark has:
- 128 GB unified memory
- support for inference on models up to ~200B parameters
- 273 GB/s memory bandwidth

the 128 GB is genuinely useful because it lets very large models fit locally on a desktop-class machine

but fitting a model doen't mean that it can run it at what I'd consider decent speeds (>=25+ tps)

I suspect memory bandwidth is going to become a much more familiar PC spec as local AI goes increasingly mainstream

[Ruthvik Dev @ruthvik_devx]

@iamAlexTurnbull definitely possible depending on the product, mainly how unique it is

Tesla is a good counterexample here. for years they famously spent essentially nothing on traditional advertising, but tbf Elon himself was an enormous distribution channel for the company

[Ruthvik Dev @ruthvik_devx]

its not even HQ'd in the sf bay area either!

[Ruthvik Dev @ruthvik_devx]

kinda wild that Black Forest Labs, which was only founded in 2024 and raised a $300M Series B at a $3.25B valuation in December, is already producing a text-to-video model that ranks above the latest models from OpenAI, Meta and ByteDance

seems it is still possible to compete with tech giants in AI, especially in less competitive domains

[Ruthvik Dev @ruthvik_devx]

@NielsRogge its pretty niche stuff

[Ruthvik Dev @ruthvik_devx]

i'd probably ask what "slow" means first

sometimes the ORM produces a bad query shape. but often the query is fine and the schema/indexes are the problem. sometimes the database isn't even the bottleneck

the useful boundary is probably ORM by default, raw SQL where profiling gives you a reason

[Ruthvik Dev @ruthvik_devx]

every new generation of AI models will render dozens if not hundreds of agent skills redundant

[Ruthvik Dev @ruthvik_devx]

speed might actually unlock more agentic use cases than another marginal jump in intelligence, at least for tasks where the models are already smart enough

a 20 minute agent becoming a 2 minute agent changes how often you use it, what you are willing to delegate to it, and whether you actively supervise it or do something else

[Ruthvik Dev @ruthvik_devx]

just a heads up, the last 10% is disproportionately hard.

replacing an editor or media player with FOSS is doable, but the difficulty is mainly in replacing Stripe, frontier AI models, Slack/Zoom and Google Workspace

basically anything that touches infrastructure or has significant network effects is hard to switch away from. local apps OTOH are much more friendly for open-source alternatives

[Ruthvik Dev @ruthvik_devx]

@nikiscevak huge miss by them. they genuinely could have become a tech powerhouse