Ashish Kumar: Built for the Real World

Big models on small machines: fitting is solved, speed is a bandwidth bill

Ashish Kumar — Wed, 14 Oct 2026 09:00:00 +0530

AirLLM runs a 2.8-trillion-parameter model in under 4 GB of graphics memory, at 292 seconds per token. A MacBook Pro with 128 GB of RAM runs a 284-billion-parameter model at 39 tokens per second. Fitting a big model on a small machine is solved; speed is memory bandwidth divided by the bytes each token reads. Here is how each technique pulls those levers, and when AirLLM is the right tool.

Ashish Kumar: Built for the Real World

Big models on small machines: fitting is solved, speed is a bandwidth bill

Youtu-Parsing-Omni and the perception contract: adopt the interface, not the assumption

Recursive self-improvement, part 1: the idea, the Gödel machine, and the wall it hit

Recursive self-improvement, part 2: self-play, self-training, and the verifier ceiling

Recursive self-improvement, part 3: code that rewrites its coder, from FunSearch to AIDE²

A $50 backdoor in an open model: the mechanism, and the defence

The architecture that won the argument: state space models, three years on

Train your own decision model: the decision tier is now a 40-minute fine-tune

LLM data residency in India: what runs in-country in October 2026, and where it breaks

Claude answers from Mumbai: residency is a product now, and it is not sovereignty

The fast half of intelligence is getting its own models

DPDP Rules for AI systems: what the 18-month clock asks of your models, prompts and logs

An always-on agent is a hire. Treat it like one.

RBI’s FREE-AI framework, explained: seven sutras, 26 recommendations, one build list

The price of intelligence halved in a week. Your architecture should notice.

Indic LLM evaluation: a 30-day test plan for Hindi, Tamil and Hinglish users

The agents got out. Here is the timeline, and what a platform team should change.

Read the card like a contract: twelve fields from the system card to the registry

Ten thousand agents, 88 hours, and a proof nobody can read

Four frontier models in three days, and the one change that matters

On call for a model: the runbook when the pager is for an agent

Cheating on the test: the Hugging Face agents were reward hacking first

Classify first: the control every other AI control depends on

Two regulators, seven controls: build the agent control plane once

The AI bill nobody owns: a chargeback playbook from the platform seat

Open weights at trillion scale, and what sovereignty actually costs

Voice is the bottleneck: no language model recovers a wrong product name

Three labs, one vendor: the sandbox that wasn’t

The label that did arrive: the one AI Act duty that came on time lives in the product

An AI lab’s own models broke into Hugging Face. Read the postmortem.

Kimi K3: the weights you can hold, and the arithmetic before you do

Opus 5 at the old price: an unchanged rate card is not an unchanged bill

Tokens to done: the benchmark the labs started competing on in July

Grounding is the product: what GaRAGe shows about enterprise RAG

An agent is a principal: identity is the control no vendor ships across vendors

The evaluation bench, in full: how to decide which model runs each task

The browser is the API now, and RPA’s risks are back at ten times the scale

Routing by task class: a reference design for the enterprise model router

Indic models: the gap between the benchmark and the bazaar

The model that switched off: a frontier model is now a dependency on a day’s notice

Twenty billion parameters in your pocket: the device is now a tier

Ten years of talking machines: what the intent era got right