Follows and tweets on AI - AI Transformation- Product Management* Public Cloud* Cyber Security*

Chennai, India
Joined May 2009
Nirmal retweeted
I have taken so many pictures of the sign, especially in winter because it's usually dark when I leave the office. Camera: RICOH GR IIIx Focal Length: 26.0 mm Aperture: f/2.8 Exposure: 1/100 ISO: 400 photos.deepaksingh.net/World… #dailyphoto
2
1
12
849
I just created a playbook to help you learn AI engineering. It gives you: • Core concepts behind modern AI systems. • Condensed notes to understand AI engineering stack. • Must know techniques to build, deploy, debug & scale AI apps. (24 hours only!) To get it for FREE: 1. Like, Retweet & Follow @systemdesignone 2. Reply "Playbook" Then I'll DM you the details.
521
365
6
730
46,283
Nirmal retweeted
some thoughts on working with ai models • context as infra • taste as config • verification for autonomy • scaling via delegation • closing the loop eugeneyan.com/writing/workin…
37
44
7
291
27,243
Nirmal retweeted
The Harness is a Context Manager on Behalf of the Model What happens when the context window fills up and who decides? This decision is external to the model - The Harness designer must have some opinion here! Decisions like this are crucial in turning a model into a great product for end users. The context window is a sacred boundary beyond which all model computation actually happens. Context engineering is important because designing what gets passed over this boundary is the main determinant of agent performance. Harness design is how you decide how this boundary gets managed. Our create_agent primitive in LangChain exposes the one of the simplest Harnesses for builders to extend, a ReAct loop with support for tools, middleware (hooks), and model choice It’s a great place to start in agent building because it forces you to think through and contend with all of the design details that transform a simple agent loop into a purpose-built agent for your tasks The first time you hit the context boundary in a simple agent loop, the API will just error out and your agent run will end. The API contract only supports a max number of tokens. A harness helps you get in line with the API contract by managing context via strategies like truncation, compaction, offloading, and targeted context eviction This is just one decision to think about in Harness Design, many more come up as you build such as agent specialization via Subagents, Tool design, Skill design, and more. Each of these are important in extending a model to make it into a useful product for users. create_agent is a great level of abstraction to start building agents. Builders can go up a level to deepagents for a more out of the box agent experience and even further to Fleet as a more out of the box product experience. Or they can go down to the runtime execution level to LangGraph like @caspar_br had talked about starting from a simple harness to build a great agent helps you learn fundamentals of how models work + good design patterns that turn them into great agents and products
8
38
4
327
21,445
📣 Shipping software with Codex without touching code. Here’s how a small team steering Codex opened and merged 1,500 pull requests to deliver a product used by hundreds of internal users with zero manual coding. openai.com/index/harness-eng…
73
195
80
2,019
572,394
New Google paper says LLM inference is limited by memory and networking, and it proposes 4 hardware shifts. Inference is separated into Prefill, which processes input in parallel, and Decode, which outputs 1 token (a small text chunk) at a time while constantly reading a Key Value cache, a running memory of past text. LLM serving speed and cost problems come from the Decode phase, where the system keeps fetching past context from memory and talking across chips, not from a lack of raw compute. The paper argues today’s data center accelerators were shaped by training needs, so they often look fast on paper but waste time during Decode because memory bandwidth, memory capacity, and interconnect latency become the real bottlenecks. The authors review today’s hardware and show a mismatch, high compute helps less when Decode gets stuck on memory bandwidth, memory capacity, and network latency. High Bandwidth Flash stacks flash to aim for about 10X more capacity for model weights, the learned numbers that stay fixed during inference, or for slow changing context, and processing near memory plus 3D stacking move compute closer to memory so less energy and time are wasted moving data around. For fast user responses, the paper also pushes low latency interconnects, meaning fewer network hops and smarter switches, so the same LLM can run on a smaller cluster that costs less and fails less. ---- Paper Link – arxiv. org/abs/2601.05047 Paper Title: "Challenges and Research Directions for LLM Inference Hardware"
34
92
12
640
66,504
Really enjoyed Matt’s keynote at #AWSreInvent today. So much innovation happening in @awscloud, and you could see it with the array of launches he unveiled. So many parts of the keynote worth watching, but will point to a few: 1/ Excited about the availability of Trainium3. Trainium2 has substantial traction, is a multi-billion-dollar revenue run rate business, has 1M+ chips in production, and 100K+ companies using it as the majority of Bedrock usage today. Trainium2 has price-performance advantages over other GPU options that are compelling, and Trainium3 will deliver at least 4.4x more compute performance, 4x greater energy efficiency, and almost 4x more memory bandwidth than Trainium2: youtu.be/q3Sb9PemsSo?si=0rVn… 2/ Worth double clicking on AgentCore, which has changed the security and scalability of deploying agents into production. AgentCore is a set of flexible building blocks that can be used in any combination developers want, and AWS added two more in Policy and Evaluations. AgentCore has a lot of momentum: youtu.be/q3Sb9PemsSo?si=0Qj1… 3/ Nova Forge is a game-changer for companies wanting to customize a frontier model with their own proprietary data. Like equipping a young person with a better knowledge foundation to keep learning, LLMs are better able to solve problems and improve if they’re trained early on with companies’ differentiated data. To do so, companies need earlier versions of the frontier model and ability to mix their own data with the model’s data. This is what Forge provides and this “open training” allows companies to develop Novellas that are their own, optimized versions of Nova they can use for their AI apps and agents. Customers have been itching for this sort of capability, and Forge is a uniquely compelling approach: youtu.be/q3Sb9PemsSo?si=S0CJ… 4/ Agents will become the primary way companies get value from AI. We have built some compelling agents for our customers in Kiro (for coding), Quick (for knowledge workers to leverage their own data, analytics, and routines), Transform (to migrate from one software source to another), and Connect (call center agents). But there are tasks customers want agents to solve more autonomously and over longer durations, and new AWS frontier agents—Kiro autonomous agent, AWS DevOps Agent, and AWS Security Agent—are exciting: youtu.be/q3Sb9PemsSo?si=R95D… 5/ Finally, I enjoyed Matt’s ending 25 launches in 10 mins, both because it was action-packed and represented so much useful delivery. The reason so many people cheered these launches is because even though they’re less sexy, they’re the meat and potatoes core infrastructure needs customers have—and with so much of the total cloud infrastructure running on top of AWS, these launches will make a lot of people’s lives easier and better every day: youtu.be/q3Sb9PemsSo?si=Gusf… Enjoy (and just day 1 of announcements today for AWS :-)!
46
91
18
643
76,962
Zoho’s business would be first to be competed away by people building their own custom software built by people using @Replit @emergentlabs and @Taskade Why pay $30/seat/month for over bundled SaaS when soon even nontech ops ppl can vibe-code a custom solution in a weekend?
646
97
426
1,559
1,765,805
Just created a complete analysis of AI infrastructure opportunities covering chip manufacturers, power companies and system integrators. I shared this analysis with my 20,000+ students. For 24 hours, it's yours for FREE. Like, RT & Comment "AI" and I'll DM it to you.
3,455
1,660
77
5,260
695,337
Just leaked 110+ plug-and-play n8n templates that multiple industries are paying me $10K+ for. Your weekend project: Pick any industry, copy-paste a workflow, start printing money. The vault includes: - Real estate lead machines (sold to 2 companies) - Automotive inventory bots - YouTube-to-TikTok content farms - E-commerce customer journey automations - SDR agents that close without you - Healthcare patient follow-ups - Legal document processors - Social media engagement bots From study agents (because fuck school) to manufacturing systems - every industry that prints money is covered. The difference: These aren't broken templates with outdated nodes. They're battle-tested, profitable, and actually work in 2025. Half sold for serious cash. Half are experiments that became goldmines. All are ready to deploy. Follow, RT + Comment "vault" and I'll drop the entire collection before my accountant kills me for giving away the farm. (Sorry to every consultant charging $10K for basic automations)
5,949
2,901
68
6,967
1,020,316
How Docker Works Internally. Explained Simply 👇 47K+ read my TechOps Examples newsletter: techopsexamples.com/subscrib… What do we cover: DevOps, Cloud, Kubernetes, IaC, GitOps, MLOps 🔁 Consider a Repost if this is helpful
6
102
1
552
59,342
Wow, Kamal Haasan @ikamalhaasan knows about Replit, me, and our story and promotes it to a large audience. I'm so honored!
166
832
116
8,157
612,241
SaaS is being dismantled as we speak! We're witnessing the slow-motion collapse of an entire business model that dominated tech for two decades. The $1.3 trillion SaaS is being quietly hollowed out from within by AI agents. Here's how I see it playing out: Phase 1 (Now): AI as co-pilot. We're seeing this everywhere, Copilot for developers, Gamma for presentations, Harvey for legal research etc. These AI layers sit atop existing software, making it more efficient. The SaaS companies feel safe, even excited, as AI seems to make their products more valuable. They're bringing knives to what they think is a knife fight. Phase 2 (Next 12-18 months): The agent invasion. AI moves from co-pilot to autonomous operator. They're replacement workers that can fully operate existing software on your behalf. The dam breaks when someone can say "analyze our Q2 performance" rather than clicking through Tableau, or "optimize our ad campaigns" instead of navigating Meta's ad manager. The expertise previously bundled with the software gets unbundled by agents. Phase 3 (2-3 years): Software invisibility. The final phase happens when the agents bypass the human interfaces altogether. Why render dashboards, buttons and menus when AI can just access the APIs directly? The value proposition of SaaS, bundling software, workflow, and expertise into user-friendly interfaces unravels completely. The interfaces were designed for humans, but agents don't need them. Most SaaS incumbents don't see it coming because this isn't a classic disruption pattern. It's not about competing products with better features. It's about the evaporation of the core assumption that humans will operate software. What's more, the barrier to creating custom, internal software is collapsing simultaneously. Companies that once had to choose between expensive custom development or off-the-shelf SaaS can now spin up bespoke solutions in days instead of months. Why pay Hubspot $1,500/month for a CRM when your team can build 'HubspotForUs' with an AI coding assistant over a weekend? The same features, perfectly tailored to your workflow, with no ongoing subscription costs. This democratization of software creation means every company becomes a potential software producer rather than just a consumer. The specialized knowledge that SaaS companies monopolized is now available to anyone with access to an AI coding agent and domain expertise. It went from $1M to build an MVP to build a SaaS to basically free overnight. I bet the metrics will be puzzling at first, DAUs remain strong while feature usage mysteriously declines. The power users who drive revenue suddenly need fewer seats. Customer success calls shift from "how do I use this feature?" to "can your software work with my AI agent?" Or worse: "we built our own version that better fits our workflow." The survivors won't be those with the best features or even those who add AI features fastest (from no AI to "ai-assisted"). The winners will be companies that expose their software's capabilities through agent-friendly APIs and position themselves as the most trustworthy information sources and execution engines in their domain. There's also the shift from monthly subscriptions to outcome based software (pay per outcome, pay per task etc) but that's a tweet for another day! The $1T question: Will Microsoft, Atlassian, Adobe etc. successfully navigate this transition, or will they be the Digital Equipment Corporation of our era too invested in the previous paradigm to adapt to the new one? All I know is this will be a golden era for startups in the space. SaaS is being dismantled, piece by piece, workflow by workflow, interface by interface. Am I wrong?
565
793
227
5,816
1,121,235
This man literally lays out a step-by-step formula for a 100x return.
32
320
3
4,216
446,915