←
Ankit Prateek7,876 posts
Ankit Prateek avatar

Ankit Prateek

@ankit_appy

Reader. Geek. Gamer.

◷ Joined November 2009▧ 368 Photos and videos
Ankit Prateek avatar
Ankit Prateek@ankit_appy·Nov 3

Exactly why the real-world test matters alongside benchmark scores.

#qwen-3.8-flash-next vs #glm-5.3, running on an #RTX5090. GLM performed far better in this real-world test than its benchmark score suggested. It delivered a stunning working page first try; Qwen needed a repair round.

OffByQuant avatar
⟳ Ankit Prateek retweeted
OffByQuant@OffByQuant·Oct 27

Crouching VRAM, Sleeping Dragon: How the Strata Family Makes Massive AI Models Run by Waking Only What They Need uncategorized.blog/posts/cro… #strata #glm #localLLM #inference #nvidia #rtx5090

c0c0n2026 avatar
⟳ Ankit Prateek retweeted
c0c0n2026@_c0c0n_·Oct 9

Live Now | Supply Chain Security Village at c0c0n 2026 The Supply Chain Security Village is live at c0c0n 2026, bringing focused discussions and hands-on exploration of supply chain security to the conference. Join the community at the village and be part of the learning, collaboration and knowledge sharing happening at c0c0n XIX. 09–10 October 2026 Grand Hyatt, Kochi, Kerala Connect. Collaborate. Contribute. #c0c0n2026 #c0c0n #SupplyChainSecurity #SecurityVillage #CyberSecurity #InformationSecurity

Post attachment
Ankit Prateek avatar
Ankit Prateek@ankit_appy·Sep 29

When the weekly token-limit line starts looking less like a usage meter and more like your lifeline, you know you have #WLDS 😭

Post attachment

I think I’ve developed a new mental health condition: Weekly Limit Doom Syndrome #WLDS. At 80% usage remaining, every bug feels solvable. At 5%, every minor issue feels like an architectural dead end and the model suddenly sounds disappointed in me. (1/2)

Ankit Prateek avatar
Ankit Prateek@ankit_appy·Sep 29

Every subagent spawned at 80%+ weekly limit feels optimistic: “We are so building this.” At 10%, every new subagent spawned to diagnose the problem increases the heartbeat and anxiety exponentially. Hands start shaking. The final stage of WLDS: "..to Esc or not to Esc."

Ankit Prateek avatar
Ankit Prateek@ankit_appy·Sep 29

Symptoms: 90% left: “We’ll figure this out.” 20% left: “This may require some restructuring.” 3% left: “The project has fundamental flaws.” Complete remission immediately after the weekly reset. (2/2)

Ankit Prateek avatar
Ankit Prateek@ankit_appy·Sep 29

I think I’ve developed a new mental health condition: Weekly Limit Doom Syndrome #WLDS. At 80% usage remaining, every bug feels solvable. At 5%, every minor issue feels like an architectural dead end and the model suddenly sounds disappointed in me. (1/2)

MAX avatar
⟳ Ankit Prateek retweeted
MAX@maxxmalist·Sep 22

The race for AGI Script: Sherpa by Pocket FM Video: Seedance 2.5

1:02
Ankit Prateek avatar
Ankit Prateek@ankit_appy·Sep 10

Stream the weights, shrink the KV cache, spare the Mac 😍

💾 Smaller KV cache. Bigger savings. Compared with the previous generation, V4.1-Flash’s KV cache needs just: 🔹 1/4 the HBM 🔹 1/8 the SSD storage Cache-hit charges often account for a large share of agent costs. Compressing the cache cuts those costs significantly. 3/6

Post attachment
Ankit Prateek avatar
Ankit Prateek@ankit_appy·Aug 21

#AIME 2025, run locally, thinking on: 90.0% pass@1. Two of three misses were 64K truncations, not wrong answers – 96.7% with a 128K budget. On a desktop Mac.

Post attachment
Ankit Prateek avatar
Ankit Prateek@ankit_appy·Aug 21

Same model, Q4, #Mac vs #RTX5090: capability identical within noise on real-CVE detection and HumanEval+. The gap is physics – 2× on decode, 16× on prefill.

Post attachment