←
Hasgeek9,447 posts
Hasgeek avatar

Hasgeek

@hasgeek

Follow us to geek out on all things hardware & software ☁

◷ Joined September 2010▧ 771 Photos and videos
Shraddha Agrawal avatar
⟳ Hasgeek retweeted
Shraddha Agrawal@shraddhaag·Nov 16

wrote an article after a while! 😁 spent the last few weeks researching how cgo works, going through the Go source code. Give it a read and let me know your thoughts? great collaborating with @eatonphil! ✨

Welcoming our latest writer on The Consensus, @shraddhaag, who takes a look at some of the weird things going on when cgo checks your arguments. theconsensus.dev/p/2026/11/1…

Post attachment
Anant Shrivastava avatar
⟳ Hasgeek retweeted
Anant Shrivastava@anantshri·Nov 12

AI-BoM is becoming a bigger discussion topic and we wanted to use @rootconf with specific focus on AI to create a space to have that discussion. I will be facilitating the BOF along with @viveksahu_26 Details : hasgeek.com/rootconf/2026/su… - AIBoM is a compliance requirement - AIBoM is also a technical control that can help us with inventory - and many more other usecases. However as much as its prospects are, there is more confusion then clarity around this. Much of it comes from contrasting details. so this BoF hopes to bring people together to discuss and identify common interests and allow like minded folks a space to keep exploring.

Post attachment
Rootconf avatar
⟳ Hasgeek retweeted
Rootconf@rootconf·Nov 7

1/4 An agent that works in a demo still needs a platform that works under load. At Rootconf 2026, @phonepe and @razorpay share two perspectives: scaling AI-assisted code review, and operating long-running agents across tenants. 🧵

Rootconf avatar
⟳ Hasgeek retweeted
Rootconf@rootconf·Nov 7

2/4 More generated code. A growing review bottleneck. @tushar_knaik, Software Architect at PhonePe, shares how Arnold evolved beyond a simple AI loop into an async, distributed code-review pipeline, with humans still in the loop.

Post attachment
Rootconf avatar
⟳ Hasgeek retweeted
Rootconf@rootconf·Nov 10

1/4 AI can shorten infrastructure work. How do you check the result? Two Rootconf talks connect Kubernetes upgrade preparation with incident investigation: Adithya Parthasarthy from @bookingcom and Hemant Kumar from @One2NC

Rootconf avatar
⟳ Hasgeek retweeted
Rootconf@rootconf·Nov 10

2/4 3 weeks to 3 hours: Adithya Parthasarthy shares @bookingcom's AI co-pilot for Kubernetes upgrade preparation. Changelogs, dependencies and compatibility checks across 50+ repos and 1,000+ services, with human review in the loop.

Post attachment
The Fifth Elephant avatar
⟳ Hasgeek retweeted
The Fifth Elephant@fifthel·Nov 6

4/5 Vivek’s A100 experiments show 1.38× faster decoding at matched accuracy. Put the index back in GPU memory, and the advantage disappears. @AloshDenny asks whether training is doing the intended work. Vivek measures what limits inference, and where an optimisation helps.

Post attachment
The Fifth Elephant avatar
⟳ Hasgeek retweeted
The Fifth Elephant@fifthel·Nov 6

3/5 At a million tokens, what limits decoding when the KV cache moves to host RAM? @vivekkalyanaran examines the index scan crossing PCIe, and tests whether reading fewer bits can make the decode step faster.

Post attachment
Rootconf avatar
⟳ Hasgeek retweeted
Rootconf@rootconf·Nov 6

2/5 48 GPU hours. Just 10% of the intended data. Alosh Denny traces an LLM training run through shared-state failures and missing checkpoints, and shares the checks he now runs before committing compute. What would tell you that a running job is doing the wrong work?

Rootconf avatar
⟳ Hasgeek retweeted
Rootconf@rootconf·Nov 6

1/5 Where is your GPU time going? A training run repeatedly processes only a tenth of its dataset. A long-context inference step spends its time moving data. Two ideas/projects at @rootconf examine what happens when you measure the work your system actually does. 🧵

Rootconf avatar
⟳ Hasgeek retweeted
Rootconf@rootconf·Nov 5

2/3 20 Nov: Platforms for AI Private AI, agent scale and security, evaluation, and reliable model serving. 21 Nov: AI for platforms Code review, SRE agents, Kubernetes upgrades and query performance. Implementations, failures and engineering trade-offs.

Rootconf avatar
⟳ Hasgeek retweeted
Rootconf@rootconf·Nov 5

1/3 New dates, talks selected. Rootconf 2026 moves to 20 & 21 November. 64 submissions, reviewed and developed with community editors and reviewers. Two days examining how we build platforms for AI, and use AI to operate platforms. 🧵

Rootconf avatar
⟳ Hasgeek retweeted
Rootconf@rootconf·Nov 2

In 2025, @anantshri @adityapatawari @yudocaa & @bnchandrapal reviewed Drove's utility & architecture. Read on: has.gy/TlZP

Drove Container Orchestrator 1.34.1 is out!! phonepe.github.io/drove-orch… Discord: discord.gg/HzNcHZgcKz

Post attachment