Nikhil Verma
nikhilv.bsky.social
Nikhil Verma
@nikhilv.bsky.social
CTO at @nonfx_company, previously Ollion @BumbleEng @opera @aol

I love the web and I work to bring functionality to the web people don’t consider
Like a moron I decided to dust my computer with a blower. Ended up shorting and bricking the motherboard. Mercury is in gatorade or something
July 15, 2025 at 3:39 PM
Hey @dachsjaeger.bsky.social have you tried dxvk-async? I know in DF directs you have mentioned that some of these are placebo. But it seriously improved the jank I faced in Sekiro on my low end gaming PC. I hope you can give it a shot.
May 16, 2025 at 9:44 AM
The workflow I personally like for development is using VSCode Copilot, open the files I know are right for the context, then select "Open Editors" as the model context, then let Gemini 2.5 or Sonnet 3.7 take the wheel. It's a good balance between vibe coding and whatever we did before AI came along
April 16, 2025 at 1:12 PM
I wonder if someone has tried to open their fine-tuning data to the internet in the hopes that the AI bots crawl it and you end up getting the fine-tuned model for free 🤪
April 15, 2025 at 6:43 AM
My fave pipeline: ChatGPT for image generation, then Automatic1111 to upscale 4x. ChatGPT’s autoregressive magic nails prompt accuracy, while SD’s detailing shines in the upscale.
April 6, 2025 at 6:08 AM
I am having a lot of fun with smaller LLMs. They run cheap and fast. And if you break a problem in small chunks they do a pretty good job with 1/16 the cost
March 5, 2025 at 3:17 AM
When writing code becomes easy, being able to read and judge if the code serves it’s need becomes very important
March 2, 2025 at 3:33 PM
When you work in a team that iterates rapidly
March 2, 2025 at 7:40 AM
The "Next Edit" mode in VSCode Copilot is almost magical. Somehow it's always figured out what I want to do next. There are random mistakes sure, but the time saved overall is huge.
February 7, 2025 at 8:09 AM
Persona 5 has the greatest soundtrack in video game history
January 25, 2025 at 6:21 AM
Claude 3.5 from @AnthropicAI has this weird ability to know what the user is asking for and expects. It's lacking in gpt-4o. Infact I am using Claude responses as a few-shot prompts for gpt-4o so it can behave the way I want it to.
January 7, 2025 at 6:21 AM
Claude Sonnet 3.5 is still unbeatable in real world programming scenarios. Almost every single time it does what I ask and it runs without fail and the code is well explained and documented. Whatever they have cooked is incredible!
December 31, 2024 at 5:06 AM
This is such a good article. It captures something I've struggled to explain before.

minds.md/zakirullin/c...
Cognitive load is what matters
There are so many buzzwords and best practices out there, but let's focus on something more fundamental. What matters is the amount of confusion developers feel when going through the code.
minds.md
December 26, 2024 at 8:27 AM
I love Ted the Caver. It’s like Blair witch for blogs.

If you like horror you must read it!

www.angelfire.com/trek/caver/p...
Ted's Caving Page, with the story of his discovery in a local cave.
www.angelfire.com
December 24, 2024 at 9:33 AM
@dachsjaeger.bsky.social @dark1x.bsky.social would love to see you guys cover or discuss in df direct the Prey mod which adds native HDR DLSS etc www.nexusmods.com/prey2017/mod...
Prey Remastered - Luma
Luma is a Prey (2017) + Mooncrash DLC mod that re-writes the game's late rendering and post processing phases to improve the game's look without drifting from the artistic vision
www.nexusmods.com
December 19, 2024 at 1:21 PM
I see modern LLMs as entities with high talent and medium ability. But the cool thing is that with few shot prompts, fine tuning and agentic workflows you can bump them to be experts at specific workflows. Which is the most exciting part about working with them!
December 16, 2024 at 6:51 AM
Tried Gemini 2.0 API. Instantly ran into quota limits. No way to pay for it. 🤷
December 12, 2024 at 1:53 PM
You can get SO much mileage out of small local LLMs just by carefully crafting your few-shot prompts. They become almost as good as the 100x bigger models at certain tasks (not coding though lol Claude 3.5 is still the king there)
December 11, 2024 at 6:49 PM
Introducing starchitect.ai. We've built this over the past couple of months and it has some really cool tech that I'm proud of!

1. WASI to evaluate Regula rules on the browser. This means you can securely scan your Terraform files without having to upload them anywhere. (1/6)
Starchitect.ai
Achieve SOC 2 Technology Compliance Effortlessly, Get started with Starchitect and take the first step toward SOC 2 compliance at no cost. Scan, validate, and fix compliance issues in your Terraform p...
starchitect.ai
December 10, 2024 at 6:35 AM
Has Apple’s autofill pin code feature stopped working for anyone else?
December 2, 2024 at 7:27 AM
It’s very simple. People selling servers will say servers are the best thing ever. They will cherry pick facts, hide axes or crop timelines to suit the narrative. Nothing wrong with it. It’s their job. Our goal is to look past the marketing and see what’s best for our use case.
December 1, 2024 at 3:44 PM
Using AI to build code is exciting and frustrating at the same time. Wasted 4 hours yesterday because Claude led me down the wrong approach. Then I took a step back, used good ol pen and paper to figure out the right approach first. Then Claude was helpful somewhat.
November 29, 2024 at 6:08 AM
I want a service to extract JSON structure from PDF. Tried LLM solutions like numind but they hallucinated stuff (strict no no). Tried unstructured which does OCR but it messed up the table structure. Probably I can try more expensive solutions like Claude or OpenAI but there have to be alters?
November 27, 2024 at 12:31 PM
On one hand there is absolutely no way that life doesn’t exist elsewhere. But on the other hand these galaxies are so far from us, they might as well not exist
JWST just dropped a new photo of sombrero Galaxy
November 26, 2024 at 5:07 AM
November 23, 2024 at 9:13 AM