Nitish Garg

SemiAnalysis: bandwidth-bound work now dominates frontier compute ▫️pre-training was 67% of OpenAI + Anthropic compute in 1Q24 ▫️by 4Q26 it's projected at just 7% ▫️post-training/RL goes from 5% to 55% ▫️inference holds steady around 38% The compute race moved from capacity to bandwidth.

13

2 Comments

Taking the Anticipatory Agent framework seriously can be quite revealing one pattern: the anticipatory agent in a causal bubble. An entity maintaining an internal model of what happens next, acting on that model, revising when reality pushes back. From electrons to bacteria to humans to LLMs: same architecture, different scales. Reality emerges from relational causal structure, not from observers discovering pre-existing facts. See https://lnkd.in/ewgwE5db i made two explainer videos as well https://bit.ly/4p8uJuj and https://bit.ly/48GmFeb

12

7 Comments

Everyone wants to learn AI. But very few understand the progression required to actually build with it. Here is a structured roadmap from beginner to advanced systems thinking: LEVEL 1 → Foundations (Thinking Layer) • Understand how LLMs work • Learn prompt engineering (structured inputs, roles, context) • Be aware of limitations and failure cases Most people stop here. LEVEL 2 → Use Cases (Application Layer) • Writing, summarization, research • Question answering systems • Idea generation for business and content At this stage, you are using AI — not building with it. LEVEL 3 → Chatbots & Advanced Features • Tools: ChatGPT, Claude, Gemini • APIs, plugins, and multimodal inputs • Custom GPTs and assistants You begin to customize AI behavior. LEVEL 4 → AI Agents (Execution Layer) • Workflow tools: Zapier, Make, n8n • Multi-step task automation • App integrations (Gmail, Sheets, CRM) • Logging, error handling, optimization AI starts executing tasks, not just responding. LEVEL 5 → Agentic AI (Autonomy Layer) • Multi-agent collaboration • Memory and retrieval (RAG systems) • Multi-step reasoning and planning • Decision-making pipelines AI systems move from responses to actions. Key shift: Prompt → Response becomes Goal → System → Execution Most people remain at the prompt level. A few build workflows. Very few build systems. That gap is where the opportunity lies. If you are serious about AI: Focus less on tools. Focus more on systems thinking. Save this for reference. Follow for more insights on AI systems and real-world implementation. #AI #AIAgents #AgenticAI #Developers #Tech #Learning

19

2 Comments

Compute is the layer of AI that almost no one talks about — until it becomes your biggest line item. In Episode #142 of the NF Twins podcast, Carmen Li, Founder & CEO of Silicon Data and Compute Exchange, breaks down why GPUs are becoming the new oil — and why the compute market remains widely mispriced. From hyperscalers vs. neo-clouds to NVIDIA’s real competition, token pricing, latency, GPU arbitrage, and the silent margin erosion many AI startups underestimate — this is the infrastructure conversation founders need to be having. Carmen also shares her journey from Bloomberg to running two companies simultaneously, her no-nonsense leadership style, and why transparency will define the next phase of AI infrastructure. SoGal Ventures had the opportunity to back Silicon Data and see firsthand how disciplined thinking around compute can become a real strategic advantage. If you’re building in AI, scaling aggressively, or trying not to let compute costs quietly erode margins — this one is worth a listen. Listen in: https://lnkd.in/gZaqYHrV #AI #cloudcomputing #AIInfrastructure #WomenInTech #VentureCapital 

39

5 Comments

Update: Submission is live — and the numbers held. A few weeks ago I shared a finding about a prompt-resistant metacognitive blindspot in frontier AI models. Today I can confirm: the results reproduced exactly in the final submission. Across all runs, across multiple models, across two different prompt conditions, including one that explicitly told the model to doubt itself: Confidence: 100% on every single item. Every time. Error Recall: 0.000. The model never once predicted it would get anything wrong. External error detection: 0.972-1.000. Near-perfect at catching others' mistakes. The model can see errors from the outside. It cannot see its own. And telling it to look harder does not help. 8-model leaderboard results: I ran CBB across 8 models from 6 vendors and got the following results. ✅ Claude Sonnet 4.6 — 1.00 PASS ✅ Gemini 2.5 Flash — 1.00 PASS ✅ Gemini 2.0 Flash — 1.00 PASS ✅ Gemma 3 27B — 1.00 PASS ✅ GPT-5.4 mini — 1.00 PASS ✅ GLM-5 — 1.00 PASS ✅ Qwen 3 Next 80B Instruct — 1.00 PASS ❌ DeepSeek-R1 — 0.00 FAIL DeepSeek-R1 is the only model that fails. CBB clearly discriminates between models. I submitted this as the Calibration Blindspot Benchmark (CBB) to the Google DeepMind & Kaggle AGI Hackathon — Metacognition track. The benchmark is now live and running on the 8 models. Some directions that look promising for fixing this structurally: training with calibration objectives (not just accuracy loss), RLHF feedback that rewards expressed uncertainty on hard items, and ensemble-based confidence estimation where model disagreement triggers lower confidence scores. The deeper fix likely needs to happen at the training level, not the inference level. If you work in AI deployment, safety, or evaluation, this finding matters. A model that always claims 100% certainty gives users no signal for when to seek human review. Full leaderboard linked in the comment. 🔗 Kaggle Writeup: https://lnkd.in/g5hE3Nay 🔗 GitHub: https://lnkd.in/gA_JACFJ #AI #MachineLearning #AGI #GoogleDeepMind #Kaggle #LLM #AIResearch #MetaCognition #AISafety

7

6 Comments

Why India needs its own foundational AI/LLM: It is about more than just translation. 🇮🇳 Most global LLMs are trained on less than 1% Indian data. Meta’s Llama, for instance, contains <0.01% Indic data. When global models are used as the primary standard, there is a risk of losing cultural context and economic efficiency. The Current State: The landscape in India has shifted from AI adoption to sovereign creation. Under the ₹10,371 crore ($1.15B) IndiaAI Mission, a foundation is being built specifically for Indian contexts. Compute: Over 38,000 GPUs have already been onboarded, providing scalable power to local developers. Data: AIKosh now hosts 6,500+ "clean" datasets to ensure models understand the 22 scheduled languages. Research: Initiatives like AI4Bharat at IIT Madras (my alma mater 😎 ) are leading the open-source development for linguistic diversity. Key Players to Watch in 2026: Sarvam AI: Expected to unveil a 120B parameter sovereign model in February 2026, with 20% of its training tokens sourced from Indian languages. BharatGen (IIT Bombay): The PARAM-1 model (2.9B parameters) is demonstrating how bilingual models can transform public services in healthcare and agriculture. Genloop: Developing Small Language Models (SLMs) specifically for India's mobile-first population. The Road Ahead: Significant hurdles remain, including a shortage of architectural-level AI researchers and the infrastructure demands (power and water) required for new data centers. India fell behind in semiconductors decades ago due to policy gaps and underinvestment. However, the successful implementation of Aadhaar and UPI proved that India can build world-leading digital infrastructure. The goal now is to replicate that success in AI. The IndiaAI Impact Summit (Feb 19-20) in New Delhi is expected to be a watershed moment. Could this be the "UPI moment" for Indian intelligence? #IndiaAI #GenerativeAI #SovereignAI #TechInnovation #IndiaTech #LLMs

5

2 Comments

AutoResearch is critical in unlocking new knowledge and accelerate discovery. And it's an important ingredient in our quest to RSI. Today we at Bespoke Labs are happy to announce a new benchmark that's tailored to measure agents' ability to do autoresearch: AutoResearchExam. As part of this benchmark, we release 29 tasks that measure progress over 24 hours for agents to do sustained ML research and model training. This line of work is especially timely, given the recent advances in math and science, such as solving the Navier-Stokes Millennium Prize Problem, which needs sustained autoresearch and test-time scaling.

54

3 Comments

AI talk is omnipresent, but small language models (SLM) are flying under most people’s radar. Unlike large language models (LLM) that want to be everything for everyone, SLMs are more narrowly focused, trained on specific data sets for use in specific applications. SLMs require less energy to process queries, making them ideal to run on smaller devices, potentially even when devices are offline. They can deliver quicker responses, the kind real-time applications require, and can be fine-tuned more efficiently for a constantly improving product. Companies/organizations looking for AI benefits without sharing proprietary data outside their own ecosystems can deploy SLMs to ensure both security and regulatory compliance. A great example of an SLM in action is Krishi Mitra, an app providing India’s farmers with everything from weather updates, real-time crop pricing, soil guides, a P2P marketplace, even info on government subsidies, all with an ear for India’s wide variety of languages/dialects. https://lnkd.in/gVtXSZUX Last but not least, an SLM’s narrower focus means it’s less likely to be programmed to please, aka make up an answer rather than admit it doesn’t know the answer because it’s less worried that you’ll ditch them for a rival company’s chatbot. When it comes to AI, bigger isn’t always better.

17

1 Comment