Shivay Lamba

What an amazing day! Presented a poster at @PyTorch Conference: how we optimised the RLM paper with pre-fix caching & batched Sub calls with @vllm_project And a full house for our Pytorch Conference talk on running LLMs using Executorch on Android!

@Jordynww long time since OSS Vancouver!

Mutuals 🥹🥹🥹🥹 I’ve missed yall. I love yall. What’s everyone been up to??

@_nancychauhan Hello!

Hi Mutuals 👋

@mritunjay394 yes!

So I think I can be active again on twitter, now that mutuals can see my tweets again?

@facundozurdo @Beccalytics @g2i_co @ReactMiamiConf @AIEMiami Huge congratulations

🚀 New chapter! I'm joining forces with @Beccalytics at @g2i_co to help build amazing conferences like @ReactMiamiConf and @AIEMiami, and more cool things to come! Excited to create spaces, projects and initiatives to help builders and frontier labs share what they're doing 🙌

@HaimantikaM @piyushgarg_dev This was a nice conversation , 😄

Shot a quick video with @piyushgarg_dev today on what’s going to be dead next! Drop your guesses below.

@vipulgupta2048 @HaimantikaM @ialimustufa @knowShubhangi @vedantstwt @VirmaniAkshat So cool! Major fomo

One from the friends table at Google IO Connect.

@AsDivyansh @nikitabier Hey there

. @nikitabier deserves a praise for the recent algo changes. Thank you sooo much!! X is much usable now. Also, hi👋 to my friends who are seeing this.

@joeyousss so relatable!

So, I have to apply sunscreen, apply makeup, drink 2L of water, hit the gym, work and excel in it, sleep 8 hours, upskill daily coz AI is gonna take over , read books and watch podcasts that make sense - whilst also finding time to meet friends and family ???

Submissions are open for OpenAI Build Week. openai.com/build-week/ x.com/i/broadcasts/1…

@Astrodevil_ @nebiustf @MiniMax_AI @deepseek_ai pretty cool!

After GLM-5.2 vs Kimi k-2.7 Code on @nebiustf I put @MiniMax_AI M3 and @deepseek_ai v4-Pro through the same test with 3 modes: Game, Design, and Code. Rules: each model builds, reviews what it built, gets 3 attempts to fix errors. Both struggled with game generation, but DeepSeek was worse. ✅ Game Mode: → MiniMax M3: 21,998 tokens, $0.014, 27s, 1 repair → DeepSeek V4 Pro: 9,499 tokens, $0.024, 53s MiniMax at least produced something with working behavior and better structure. ✅ Design Mode: → MiniMax M3: 10,680 tokens, $0.008, 17s → DeepSeek V4 Pro: 8,685 tokens, $0.022, 46s Both were decent. MiniMax was faster, cheaper, and felt cleaner. DeepSeek leaned more generic/AI-slop. ✅ Code Mode: → MiniMax M3: 34,705 tokens, $0.022, 40s, 1 repair → DeepSeek V4 Pro: 8,335 tokens, $0.021, 58s MiniMax produced much better usability. DeepSeek broke responsiveness and the app didn't feel coherent. Second run, same pattern → MiniMax usable, DeepSeek unresponsive. Overall: MiniMax M3 is stronger for practical UI/app generation. DeepSeek V4 Pro may be token-efficient, but struggled with usable frontend output in these tests. Pick based on your workload. 🔥

@adocomplete huge congrats!

Career update: I'm evolving into a Member of Technical Staff at Anthropic 🚀 Turns out if you hang around brilliant builders long enough, you join them.

@HaimantikaM very helpful

If you work in product based companies, optimise for these things (it will help you in delivering better at work + also learn better): - Dogfood as much as you can. Use the features you are building, your friends are building. Build use-cases, give feedback, learn how things are built. - Work cross-teams. Even if your work isn’t something that requires a lot of collaboration, talk to people, setup coffee chats, offer to help. - Document everything. Document your work, help the marketing/content teams with content. Document ideas. Document feedback. Build a brag doc. Write EVERYTHING.

@ishandutta0098 @Adobe huge congrats!

Career Update: Promoted to MLE III at @Adobe!

I am subscribed to Maintainer Brief to get weekly updates about @mesheryio You can also get daily/weekly updates around your open source projects you maintain using the tool 🫡

@the_joshua_hill Let's gooo!!! Super excited to see this

Some teaser results for a new quantization method we've been cooking up🧑‍🍳 GLM 5.2 is getting even faster

@vadiamit Would have loved to meet you, but didn't get an invite this time 🥲

yesterday set the bar. next up: I/O Connect Bengaluru. if you’re building, come say hi

@brianbellx @rudrakshkarpe might be interesting for you

I removed 423 GB from GLM‑5.2 without changing the model. 1,403 GB → 980 GB. 753B weights. Bit for bit exact. No quantization or retraining. The weights remain compressed in VRAM instead of rebuilding the full model first. Full writeup and repo in the next post.

@aaishika @diplosrunclub Let's gooo

Participated in two @diplosrunclub events last year. Upping the game by committing to four in 2026. 🙇 Denver ✅

@HaimantikaM @nebiustf Absolutely will test it

@HowDevelop @nebiustf Nicee! Also try to use DigitalOcean’s serverless inference - maybe for TTS🤓

Open source maintainer burnout isn't an effort problem. It's an attention problem. The signal is buried under hundreds of issues, stale PRs, and scattered HN threads. So I built Maintainer Brief, a weekly intelligence email that does the triage for you. powered by @nebiustf 🧵

Fear of missing out on Fable makes you ship quickly, but now I can work on even more features: x.com/i/status/20763…