Nikhil Kanamarla

@paddix Data infra is making a comeback into the mainstream

The Databricks coding-agent benchmark has a big takeaway: The harness is the runtime. Same model. Same codebase. Very different cost/perf. Because the harness controls: repo search context hygiene file selection tool orchestration test/repair loops state across turns @databricks showed Opus 4.8 via Pi at ~$0.74/task vs. ~$1.94/task through the native harness, with comparable quality. That delta is not model intelligence. It is execution architecture. For agents, price/token is the wrong metric. Price/completed task is better. But the real unit is: model × harness × context strategy × tools × eval loop The next wave of AI engineering will be model-routed, harness-aware, eval-driven systems optimized for completed work per dollar.

@OpenAI Yay! Excited to start using it in codex

GPT-5.6 Sol, along with Terra and Luna, will launch publicly this Thursday. We’re expanding preview access globally now.

@Sagar_Lekhraj_ Those are not my company, Linkedin just groups them together

@nk_developer1 Yeah, but there were internships on that portal.

After leaving NVIDIA, I’m hiring a few founding engineers for my stealth startup. We believe AI is creating an entirely new class of data infrastructure problems. We’re building foundational distributed systems for this new era of computing. We’ve already validated the problem with early customers and are backed by leading early-stage venture capital firms. I’m now looking for exceptional distributed systems engineers to help build the company from the ground up. Every early engineering hire will have an outsized influence on the company’s technical direction, product, and culture. Job posting: lnkd.in/gyX5gYxT I’ll be sharing more about what we’re building over the coming months. Stay tuned.

@mcdaaku Sorry, only remote within the US

@nk_developer1 Hey, are you hiring any remote peeps? I am open to work trials as well. linkedin.com/in/ankit-thawal

@Sagar_Lekhraj_ It’s not a internship

@nk_developer1 I have applied to the intern position. I would love to get experience and progressive learning.

@gortron Interesting!

S3 is the perfect place to store data, until you try to search it. Two months ago I launched Firn: open source vector + full-text search on S3. Since then: 400+ stars, Python client (pip install firn), multi-vector support, semantic caching, 7 validated storage providers. github.com/gordonmurray/f…

@KJHMiao I see. How would you define the term, though?

@nk_developer1 Definitely revolutionary MM-model but still not for code and agentic workflow. Guess MM is loaded term ha

Been saying this for years: we need our own open-source frontier model ecosystem. Surprisingly too, we don't even have an alternative to a frontier multimodal model. Kimi is the only one.

@KJHMiao What about the Cosmos family of models? arxiv.org/abs/2606.02800

@nk_developer1 Nemotron Ultra is still text-only

@jxmnop Interesting read

new paper from our work at Meta! **GPT-style language models memorize 3.6 bits per param** we compute capacity by measuring total bits memorized, using some theory from Shannon (1953) shockingly, the memorization-datasize curves look like this: ___________ / / (🧵)

@nshmadhani @Meta Done! Happy to chat 💬

@nk_developer1 Hey Nikhil, I am developer @Meta working on a very similar problem statement. Happy to connect and help build. Can you follow me back so I can dm you?

@nageswarkakolla Yeah haha, I'm so used to Linkedin

@nk_developer1 i love this job post, u come to twitter and to share linkedin job post, sorry, no offence, i just found this funny or irony of life

@willccbb Yup the classic path, it works though!

i took a fairly unconventional path into ai research. i did my undergrad in computer science, and then did my phd in computer science. i spent some time in big tech, and also in quant finance. but ultimately, i realized that my true calling was working at a neolab

@deedydas The data layer is going to be so important!

“The dirty secret in AI is that everything is a data and an eval problem. The best models have the best data and best internal benchmarks. The mid ones buy a lot of data, not the best, and hillclimb public benchmarks. (you need a lot of compute too)” – Stanford CS Professor

@jhleath All of this doesn’t matter too much if your company is 1 person ;)

@jhleath Nice article, the code review process was the last blocker for my old Nvidia team. Focusing on ensuring testing artifacts are available when putting up the PR and having a good CI/CD pipeline seems like a good step forward. Therefore the review process can become lighter?

@vchennai2 They need to improve availability, at least 2 9’s!

So this is how I'm supposed to access my repo when github goes down

@PKUWZP +1

To all “frontier AI labs” recruiters or HMs, one suggestion I have is to really look at objective qualifications, e.g. publications in top ML venues, PhD training, research expertise, and OSS contributions etc. AI research should not be like hedge funds, some random guys whose name is among a few hundreds names on some tech reports should not rank higher than solid researchers who publish top ML conference papers. Unless it’s for the senior leadership positions (which is reasonable to rely on connections), let’s filter out nepotism, connections or secret recipes etc. This field needs to focus on science and innovation, and reduce the false positives of hiring. Let’s make sure the real talents are given the opportunity to do their best work.

@jhleath I must expand our presence in northern NJ too 😂

i have long since argued that there are only two tech powerhouses in America -- San Francisco and Atlanta this move by Corgi only confirms what we already knew as a result, Archil will be moving our HQ from San Francisco to Atlanta immediately each Archil engineer will receive 1 acre of land and central A/C which is always set to 68º or lower