I like @deedydas's work but but this take misses context Sarvam-M isn’t a vanity fine-tune; it’s India’s first open-weights 24 B Indic-centric LLM built under brutal GPU & data scarcity. Judging it by few hours of HuggingFace stats badly misses the point. Most people outside India don't appreciate that compute is quite the invisible ceiling - H100 clusters are still not commercially stocked in India - US export caps tightening next week will squeeze supply even further - Indian teams literally queue for hours of A100/H100 time that US & CN labs get on tap Data is also the long-tail problem Indic languages form <0.01 % of CommonCrawl. You read that right—two orders of magnitude less than Chinese or Spanish. Any local lab must build its corpus first, then train. That’s months of ETL before the first gradient step. Synthetic data is GPU-constrained. Talent pipeline is still forming HPC + RLHF + compiler-level optimisation is new ground in India; Sarvam’s run has already up-skilled dozens of engineers who now know how to wrangle 10 k GPU-hours, FP8 PTQ & GRPO reward engines. Their detailed blog post democratizes a lot of this learning. You can’t AWS-credit your way to that muscle memory. What Sarvam actually shipped - 3.7 M high-diversity Indic prompts, deduped & quality-scored - Two-phase non-think/think alignment that adds+2 pp on IndicGen - GRPO RL with partial-credit rewards—LiveCodeBench jump 0.23→0.44 - FP8 + look-ahead decoding: 2× tokens/s, ½ $/M tok on H100 That means a 🇮🇳-hosted midsize model now matches Gemma-3 27 B and Llama-3.3 70 B on Indic reasoning while costing a fraction to serve. That’s some engineering leverage & definitely not hype. Model adoption is anyways a long-tail - one needs to ship multiple models of non-frontier quality to eventually be able to get to the one that's truly at the frontier (at least along dimensions that we care about). Plus, there's a whole host of Indic-language use-cases where this sovereign model would work much better compared to using any other open-weights model. Look at (LiveCodeBench 0.23→0.44, 2× tokens/s) If you ask for stats, you'll learn that some of their conversational AI platform reaches out to about 50M+ people in just a week. What's next possibly? - Maybe we all recognize the data problem & do Nation-scale data-collection drives (something like CommonCrawl-IN) - Public RL-as-a-service clusters so smaller labs can replicate GRPO - For devs wanting to push the Indic NLP forward, consider forking Sarvam-M, fine-tuning on your domain corpus, benchmarking on Indic-Eval, contributing back patches. Each derivative model widens the knowledge base & closes the English–Indic gap. In summary, celebrating Sarvam's work (I'm not an investor) isn't nationalism, it's recognizing an innovation feat under constraints - India can't out-GPU Mountain View today but there's technical merit on display here, regardless of the metrics. 👏 @pratykumar, @AashaySachdeva, @HarveenChadha & other friends from @SarvamAI Here's to more AI in 🇮🇳

Framework to think about post-training x Applied AI There's something important happening here that has interesting downstream effects My mental model is that once an Applied AI company owns or sits closer to a high-volume, structured but complex workflow, post-training of open weight models w/ domain specific RL + reward shaping for token efficacy becomes a rational + high-ROI move @harvey's Tenet on top of @Kimi_Moonshot's K3 base w/ long-horizon legal environments is the clearest public proof General frontier models remain necessary for many tasks but the post-training math works when 3 conditions collide today: (1) deep vertical expertise + prop or hard to synthesize workflow data (2) sufficient vol of similar long-horizon agentic tasks so that the cost of env design + RL is fixed & amortizes (3) either high-token burn at scale or task types are under represented in general pre-training (solid evals are a must for the same) I also believe that this pattern is highest leverage in knowledge work verticals that are doc-heavy, multi-step, citation sensitive & currently billed at high human hourly rates Put another way, the economic argument is that human expert labour is the dominant cost & inference is already a material OpEx line & thus, a 2-10x improvement in intelligence/token or equivalent cost reduction for a certain quality directly expands margins & enables scale that pure API consumption cannot 👏 @gabepereyra, @ItsJulioPereyra, @calvincongelado, @nikogrupen, @vtrengarajan & team

More than three years ago, a few friends and I organised a GenAI community meetup. That is where I first met @pratykumar & @vivekrag What stayed w/ me from that conversation was a view that felt fairly contrarian then, especially from India: if you wanted to do AI work that retained value, you needed foundational depth & the ability to operate across the stack. Sharing a first name with Pratyush has since led to a few pleasant mix-ups across the ecosystem. More importantly, over the last several quarters, I’ve had the opportunity to work informally but closely with Pratyush, Vivek and the wider @SarvamAI team. The closer I’ve seen them operate, the stronger my conviction has become that frontier AI can be built from India. That makes @ActivateSignal's investment in Sarvam as part of the Series B especially meaningful to me. Finding, understanding & supporting more such teams building at the frontier from India is exactly why we are building Activate. Activating AI in🇮🇳! @aakrit @soaibgrewal

Interesting way to look at it, as it compresses the capital intensity of the frontier capabilities: (1) Pure parameter or data scaling under the dense transformers regime produced the familiar power-law curves (2) The effective compute required for a given capability level fell dramatically once we shifted to sparse MoE, better expert routing, sparse attention (DSA in V4), lower-precision compute, kernels that are h/w aware (3) Training costs are no longer a function of pure FLOPs as architectural & systems progress can substitute for that raw scale (4) For me, the interesting question then becomes whether the capability frontier, measured by the best achievable performance as a function of total training compute, is still advancing at a rate consistent w/ historical scaling or whether we are mostly filling in the Pareto front of cost v/s performance for a given capability level

Several interesting things to note here: (1) @Kimi_Moonshot just invoiced a hyperscaler and that too in public. This is a new monetization mechanism at a layer everyone thought was commoditized (2) Clouds don’t do Day 0 work for models they expect nobody to use. At minimum, Google expects enough developer/enterprise pull for Chinese open weights inside GCP to justify the engineering (3) Hyperscaler license fees to open weight labs is structurally analogous to content licensing - and it converts model IP into a royalty stream that survives commoditization of serving (4) NVIDIA taxes even this transaction (B200/GB200). TPU absence from the K3 stack (for now) is a quiet negative for the TPU-as-moat narrative on 3P models (5) The ratchet is the real story. K2 (Jul '25) asked scaled deployers only for attribution - displaying "Kimi K2" in the UI. K3 (Jul '26) asks for payment. Will be fascinating to watch DeepSeek V5 and GLM 6 licenses

Directionally correct but yet incomplete Kaplan’s scaling + Chinchilla optimal scaling laws established that base model intelligence is a predictable function of training FLOPs + data vol. + param count Thus, frontier capability remains gated by massive centralised training runs However, once weights exist, inference economics decouple: quantisation, distillation, spec. decoding & optimised serving stacks (vLLM, Tensor-RT) deliver upto 50x gains For the broad middle of the capability distribution tasks, open weight models already deliver 80-95% of utility at dramatically lower costs + latency & superior control There’s a small nuance to appreciate (h/t to @bookwormengr): so long as the inference economics continue to hold, labs will continue to push the scaling curve & yet open model development will continue to persist so long as economic incentives to participants is clear (much like Linux was via h/w, service & apps)

When we wrote our first cheque to @TanayLohia1, there was no company & not even a deck What stood out wasn’t the answers he had then. It was how quickly he came back w/ a better one @BioMandrake is built on the same idea: every experiment should make the next design better @aakrit, @soaibgrewal, @ActivateSignal

I think the Higgsfield → knowledge work analogy misses an important distinction: (1) Orchestration’s product value comes from model specialisation while commoditisation improves its economics. @higgsfield operates in a particularly attractive structure: visual models are different enough that routing matters but sufficiently contestable that the application can still own the workflow (2) Creative generation is also a different domain. Model differences are more legible and failures more locally inspectable. The medium has explicit control primitives - identity, lenses, camera motion, shots, audio & composition graphs Text models overlap more, while the difficult problems sit above inference layer: persistent context, permissions, tools, verification of LHT, exception handling & recovery (3) @higgsfield’s edge therefore isn’t simply “many models.” It is the domain-specific production layer built above them: 3.1: Soul ID for trained, persistent identity 3.2: Soul Cast for reusable synthetic actors 3.3: Cinema Studio for explicit optical, camera & motion controls 3.4: Canvas for structured, reusable production graphs Seedance, Kling + Veo are specialised upstream engines. Higgsfield’s advantage is turning them into one coherent production system - a set of deterministic control surfaces around stochastic models There are LLM equivalents - memory graphs, schemas, policies, tools & agent DAGs - but they are less self-contained, distributed across organisational systems & much harder to verify end-to-end (4) That is @perplexity_ai’s structural risk. Several suppliers underneath Computer are also building horizontal knowledge work agents themselves Excellent orchestration can produce a great product w/o producing a durable control point but durability comes only if you get to own assets that model suppliers cannot easily absorb. Some potential ideas: search infra, persistent context, prop outcome loops

This is very telling about Meta's product strategy: Combining broad consumer distribution (AI thinking mode embedded in apps w/ billion DAUs) w/ aggressive low API pricing is potentially a deliberate wedge to maximize the volume & diversity of real-world agent interaction trajectories It's a recognition that high-fidelity agent trajectory data, sequences of tool calls, state changes, CUA actions, multi-step outcomes, user iterations & implicit success/failure signals, is increasingly becoming one of the scarcest & highest leverage resources for scaling agency further as & when base-model pretraining returns diminish on certain axes Would love to learn what others think

Sharp observation that forced me to think a lot. Sharing my possible hypothesis below: (1) The 2025 equilibrium - frontier model as teacher, small distilled model as workhorse - was an artifact of what the marginal query looked like. The unit of consumption then was a single turn chat w/ low quality thresholds. So latency was binding & price-per-token was the decision variable. The cheapest model above the quality threshold won & hence, you'd distill it down. But 3 forces changed that - (2.1) First, demand moved to long-horizon agentic work, where value is convex in capability. Task success compounds multiplicatively across dependent steps: at 99% vs 95% per-step reliability, a 100-step task completes 37% vs 0.6% of the time - a ~60x gap from a "small" capability delta. (2.2) Second, the correct denominator is cost/completed outcome, w/ labor in the loop - not price/token. Bigger models do more serial computation per token & small models must compensate in token space w/ longer chains and retries & that exchange rate (parameters → tokens) is poor + degrades with task complexity (2.3) Margin structure: Sparse MoE made big cheap to serve & commodity hosts price open weights near marginal cost. A closed lab's mid-tier must carry training amortization, margin + the opportunity cost of scarce h/w accelerators better allocated to frontier demand. The mid-tier gets squeezed from below by a competitive outside option & cannibalized from above by its own flagship's falling cost-per-outcome (3) Speculative but capacity economics: It's also a function of the kind of compute installation that a lab has for inference. A lab is not really selling tokens - it is selling GPU-time packaged into different SKUs. Every SKU is a wrapper around the same scarce underlying good. So the mid-tier model's price isn't just "serving cost plus margin" but also "what would this GPU-hour have earned serving frontier tokens instead?" Would love to hear from folks like @bookwormengr, @teortaxesTex, @dchaplot

I don’t think people are pricing the downstream implications of this yet frontier labs (Anthropic’s new neglected-disease program & “Claude Science” tooling) & pharma are about to pour enormous incremental compute into generative models for target discovery + molecular design Yet the binding constraints - and therefore the highest-ROI startup opportunities - sit downstream in the infrastructure & data layers: toxicology, preclinical translation, trial design, high-quality experimental data collection, regulatory-grade evidence packages, CMC & real-world evidence Pure in-silico players risk hitting a validation wall, pure wet-lab players lack the speed & generalization of modern AI. The winners likely will be companies that close tight, high-fidelity Design-Build-Test-Learn (DBTL) loops w/ physics-informed models + proprietary experimental ground truth In that vein, @TanayLohia1 & crew are building AI-native research infrastructure for programmable biology, starting w/ next-generation gene editors & de novo protein design. Their ORBIT framework + wet-lab validation engine already demonstrated generalization by winning the Adaptyv RBX1 binder competition on a problem class it does not normally target Come jam w/ us if you’re working on programming biology, there’s loads to build there from 🇮🇳

One potent way to counterposition against big labs is to own a structural, compounding asymmetry that base-model progress does not erase - (1) proprietary data flywheels generated from real production deployments (clinical notes, wet-lab assays, regulated workflows) that labs cannot access w/o equivalent trust, regulatory clearance & operational footprint (2) deep workflow embedding + switching costs inside entrenched systems (EHRs, hospital operations, scientific pipelines or sovereign stacks), regulatory/liability navigation that imposes multi-year lead times (3) In markets like India, data-residency, on-prem/HCI & national-priority alignment (DPDP, IndiaAI, PSU/healthcare preferences) that global labs have limited incentive or capability to replicate at depth. Biotech examples combine AI w/ physical wet-lab infrastructure to generate proprietary experimental data at scale that software-only players cannot match cc: @TanayLohia1, @b1shtream

This is a great point by @harshilmathur Slack was never optimized for large-scale human attention economics It was optimized for fast, threaded, searchable conversation among engineers and operators The byproduct - millions of messages carrying decisions, tradeoffs, tribal knowledge & cross-functional context - is exactly the training + inference fuel agents lack when dropped into an enterprise This is superior to most RAG setups because the context is live, conversational & implicitly prioritized by human attention patterns Humans optimize for signal-to-noise and low cognitive load; agents thrive on exhaustive context because they lack the lived experience and institutional priors that humans carry implicitly This inverts the traditional collaboration tool design principle & positions Slack (under Salesforce) as the de facto “office OS” for the agentic era - the place where humans & agents co-work, not merely co-exist @jsensarma would have an interesting perspective from his vantage point

Today’s dominant stacks (Python + PyTorch + generic CUDA libs + serving frameworks) carry a heavy “abstraction tax” in interpreter overhead, indirection & suboptimal memory/comms scheduling This is acceptable for research velocity but punitive when every cycle, byte of HBM bandwidth & NVLink lane must be maximized on bandwidth-bound hardware like GB300 By going low-level w/ aerospace-grade determinism & optimization discipline, we could push utilization dramatically higher, cut latency/throughput costs & better support compute-intensive reasoning, test-time scaling + agentic workloads that GB300 itself was designed to accelerate (higher FP4 density, 2× attention perf, vastly larger per-GPU HBM) On the positive side, this could meaningfully improve Grok’s unit economics (lower $/token or higher queries-per-GPU), user experience on X (faster responses, longer coherent contexts) & SpaceX AI’s ability to scale reasoning models w/o proportional hardware spend Power user of Grok 4.3, look forward to Grok 4.5