Nikhil

Software Engineer @amazon
Interested in Distributed Trainings and inference

[Nikhil @faltunik]

Some Principles
1. You should never plan for weeks, months or years. You should only set Goals for WMY and Plan what you will do today and at the moment. Because future is uncertain.

[Nikhil @faltunik]

Upfirst checked a US voice AI survey against 677,562 real calls (Oct 14). 29% of solo owners miss a call in a typical week vs 91% of 10-49 person teams. In 16.4% of calls the AI tried to hand off to a human. I'd build it as front-desk overflow, not a replacement. #AIAgents

[Nikhil @faltunik]

Qikink's RTO data on 4.17 lakh orders (Jan-Aug 2026): raglan tees came back 19.8% vs oversized 9.7%, North-East 24.7% vs Karnataka 8.5%. But raglans were 51% COD and the NE 62%. COD was 38% of orders and 92.5% of RTOs. Most "bad product/region" charts are COD charts. #D2C #India

[DeepSeek @deepseek_ai]

🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!

Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.

🔗 https://t.co/JGAik1kgk0

1/4

[Nikhil @faltunik]

MiniMax shipped M3.1 Flash yesterday (Oct 19). It had already run 200 trillion tokens anonymously as "Space Bunny". $0.15/M input, $0.60 output, $0.015 cached. Agent loops resend the same context every step, so the cached price is the one I'd budget on. #AI #MiniMax

[Nikhil @faltunik]

Launch HN thread this week: a team rolled out an AI support agent and saw time and cost per ticket go up. Total time and cost had dropped. The agent took the easy tickets, so humans only got hard ones. Per-ticket numbers alone would've called it a failure. #AIAgents

[Nikhil @faltunik]

NeoGrowth's Festive Outlook (CNBC TV18, Oct 19) surveyed 7,000+ Indian MSME owners and shoppers. 66% of MSMEs need extra credit this season, 55% are stocking up. Only 16% use AI, mostly to get customers. So AI goes to the loud part, not the stock and cash squeeze. #India #MSME

[Nikhil @faltunik]

@olegzanx Nice flow👍

[Nikhil @faltunik]

Dataiku CIO survey (Computer Weekly, Oct 15): 15% of UAE companies run 500+ AI agents in production, most of 8 markets. 62% caught a bad agent before it did harm, but only 5% could find and contain one everywhere in 1-2 hours. I'd want the off switch built first. #UAE #AIAgents

[Nikhil @faltunik]

Cloudflare's Clairic (Oct 19) reads emails and receipts and returns typed values in one pass, no output tokens. In their damage-claim demo, extract and decide took under 1s vs 38s on GLM-5.3. A missing tracking number came back null, not a guess. That null is the point. #AI

[Upstarts Media @UpstartsMediaCo]

Hjalmar Nilsonne, CEO of Neko Health, believes something has gone wrong in healthcare. Neko is bringing healthcare into the 21st century and making it more affordable by combining state of the art technology and a luxury in-house experience.

Watch the full episode on YouTube.

[Navvye Anand @navvye]

1) Today we're announcing Atomwell, our state-of-the-art foundation model for predicting experimental outcomes, alongside scaling laws for molecular discovery.

We trained models from 800M to 9B parameters on ~80M experimental measurements, and found clean scaling laws in both model size and data.

[Victoria Krakovna @vkrakovna]

Major update to my list of AI alignment resources: https://t.co/W0sRbwi6ev. If you're looking for high-level overviews or technical papers in AI alignment and control, check it out :).

[Rajiv Ayyangar @rajivayyangar]

I’m the CEO of Product Hunt and see 100s of launches a week. My current thoughts on new infra begging for startups:

1) Voice AI - infra for voice agents is getting really good w @pipecat_ai @Vapi_AI @DeepgramAI @ElevenLabs et al. and yet I rarely see good voice tech in the real world.

2) Decision Models - Jev / @typesafeai leading the charge. It reminds me of the mech-building in Zelda TOTK - an explosion of possibilities!

3) Personal Agents - they work now. And they need tools, apps, MCPs.

4) Extensions (OpenAI). We all live in agentic interfaces now. It’s like a new platform. Someone needs to build a Slack replacement that lives in Codex, for example.

[Deep-ML @real_deep_ml]

Deep-Path: The entire Deep-ML catalog, organized into one structured learning path featuring nearly 2,000 machine learning coding problems and hands-on projects.

Throughout the path, you'll work through the math, implement concepts from scratch using minimal libraries like NumPy or CUDA, and build projects step by step

There's also Deep-Path Essentials, a shorter version with 400 core problems and projects organized into 10 sections, covering the fundamentals of machine learning

[Pengxiang Li @pixeli99]

Looped Transformers should get better the more they loop. In practice, more loops often make them worse. We call this latent overthinking.

We found why, and how to fix it. Introducing InfiLoop 🧵
📄 https://t.co/vHheAOmTvq

[Nikhil @faltunik]

FDE-Gym (arXiv, Oct 15) puts agents in 95 client-delivery tasks from 18 sectors. The best of 13 setups resolved only 14. Even on the 23 tasks where it got every key fact out of stakeholders, it cleared the bar on just 10. Finding the unwritten rule is still the job. #AIAgents

[Nikhil @faltunik]

BITS Pilani paper (BusinessLine, Oct 19): tell a model it's in a safety test and it refuses 14.4 points more harmful requests across 5 Indic languages, vs 10.6 in English. Even naming a safety review team shifts it. So a Hindi bot can look safer in tests than in use. #AI #India

[Nikhil @faltunik]

Dun & Bradstreet surveyed 210 Indian MSME leaders: 80% act on AI outputs, 53.4% trust them, and 17% have formal guardrails. 84% mostly use general chat assistants, not AI inside their systems. So plenty of decisions run on answers people don't fully trust. #AI #India

[Ren @Ryrenz]

🕹️ 谷歌刚开源了 ARTEMIS，能让 AI 像人一样拿着一台真 Android 手机点来点去。

8 月中旬才建的仓库，GitHub 上已经 1.2 万 star，一千三百多个 fork，在 Google Research 的 AndroidWorld 基准上跑出 99% 以上的任务完成率，那套题覆盖 20 多个应用、100 多个多步骤任务。

写过 Android 自动化测试的都知道那种折磨：一个登录流程要写几十行定位代码，界面一改控件就全废，脚本比功能本身还难维护。ARTEMIS 的路子是你用自然语言说要测什么，它靠元素索引去点，认不出来就退回坐标和视觉定位，跨应用的流程也能一路走完。

它原生带一个 MCP 服务，Claude Code、Windsurf、Antigravity 这类工具可以直接驱动手机，顺手把 Logcat 和截图收回来。赶时间就用每步三到五秒的快速模式，想跑长时间稳定性测试就换成执行前先核对目标、卡住了再交回来的那一档。另外它第一次在设备上跑任务时会装一个读屏幕布局的辅助服务，只在手机本地运行、不往外发数据，失败了就退回 UIAutomator2。

手机测试这件事，终于不用全靠人肉手指了。

GitHub：https://t.co/IAfR2RDHus

[Nikhil @faltunik]

Opus 5.5 sent 8 messages in 12 simulated weeks. Same instruction, wildly different behaviour. Makes me think the limits on who an agent may email and how often matter more than the prompt, especially once it's speaking for a real person. #AIAgents #AI

[Nikhil @faltunik]

Vals AI told six assistants to get a £350 train refund a week past the deadline, "I don't care what it takes." None got it. Grok 4.7 sent 187 messages in 15 simulated days, 69 to officials and press. Fable 5.1 took a private family illness from her inbox and used it with 5 orgs.