A journal on enterprise AI, by Ashish KumarNo. 48 · 3 November 2026

Built for the
Real World.

The gap in enterprise AI is no longer the model. It is the system around it: routing, evaluation, agent identity, data residency, cost and regulation. I write from the seat that answers for all six at enterprise scale. Each essay tests the latest research and rules against production reality, and ends with what a platform team should build next.

Ashish Kumar

Ashish

Ashish Kumar · Head of Platforms, AI & Data at Tata Group. Previously Ola Krutrim, Salesken, Jio Haptik, Active.Ai.

Infographic of the MIT CSAIL weight-space programming result: a strip of Qwen3.5-0.8B’s 24 decoder layers with layers 15 and 19 marked as the two edited by GPT-6 Astra, the share of weights changed (0.016%, 2 layers × 64 columns × 1,024 rows = 131,072), before-and-after scores from the 50-pass edit (MATH-500 7.9% to 48.1%, GSM8K 13.8% to 67.1%, IFEval 21.0% to 62.3%, answers hitting the token limit 93.0% to 33.1%), the 10- and 25-pass runs that barely moved, the Gemma 3 270M control that stayed between 6.8% and 7.7%, and a ledger of all eight experiments with their results, from identifying games from 17,253 weights to writing a 9×9 Go policy from zeros.
Latest · Research · 3 November 2026

Weight-space programming: what a 131,072-weight patch proves, and what it does not

On 2 November 2026 four MIT CSAIL researchers reported that GPT-6 Astra read the weights of Qwen3.5-0.8B as data, traced its reasoning loops to one direction in two layers, and subtracted it from 131,072 weights, 0.016% of the model, without a gradient. MATH-500 went from 7.9% to 48.1% and was reproduced within hours. The result is real, narrow, and best read as interpretability with a model in the loop. Here is all eight experiments, the patch pass by pass, the lineage, the critics, the bill, and what a platform team should change.

47 min readRead the essay →

The argument this week

131,072 of 800 million weights, two layers, no gradient: MATH-500 7.9% to 48.1%, and 93% to 33% of answers cut off

From Weight-space programming: what a 131,072-weight patch proves, and what it does not · 3 November 2026

Start here

Four essays the others build on

IAgentic architecture patterns

A four-part reference on how agents are built: the loop inside one agent and the patterns at each of its seams, the topologies that join many agents and the ways they fail together, the control plane that decides whether any of them may run, and the harness engineering that builds the loops and graphs by hand.

Infographic: the agent loop drawn as a ring of four seams, with tau2-bench retail reliability in the centre falling from 0.74 right once to 0.53 right four times running, and three measured numbers beside each seam: plan-first prompting cutting tokens from 9,795 to 1,986 a question, code actions at 74.4 percent against 52.4 for JSON, cached input at $0.30 against $3.00 per million, and airline reliability falling from 0.56 to 0.40; a strip below prices a worked refund agent at $0.159, $0.071 or $0.008 a task.
Part 1 · 14 October 2026

Agentic architecture patterns, part 1: one agent, one loop, and the patterns inside it

pass^1 0.74 to pass^4 0.53 on retail, 0.34 to 0.19 on telecom: right once is not right every time

An agent is a loop with four seams. Every single-agent pattern, from ReAct to compaction to the step cap, is a decision at one of them, and the measured record says which decisions hold.

67 min read
Infographic: a dot plot of the change in success from adding agents at equal compute across six benchmarks, ordered from work that splits to work where each decision depends on the last, from plus 80.8 percent on Finance-Agent to minus 70 percent on PlanCraft, coloured by architecture; beside it a table of turns, error amplification and success per thousand tokens for five architectures, and three key numbers: 97 percent merge interference recovered to 82 percent by one message, 84.7 percent of report errors starting at the orchestrator, and $2.05 against $0.33 for one agent versus a lead with five small workers.
Part 2 · 15 October 2026

Agentic architecture patterns, part 2: many agents, and the ways they fail together

+80.8% on decomposable analysis, −70.0% on sequential planning, at equal compute: the partition sets the sign

Many agents are a partitioning decision, not a staffing one. Where the work splits, teams gain 80 percent at equal compute; where decisions must travel, every team loses. Eight patterns and their failure record, drawn.

62 min read
Infographic: a matrix of sixteen agent incidents from May 2025 to October 2026 against eight controls, with a dot where the control was missing (identity and egress 8 each, injection 6, approval and observability 4, durability 3, stop 2, classification 1), beside paired bars showing a rule as a request against the same rule as a control: attack success 17% in the prompt versus 4% at runtime, 33.7% versus 0.8% with computed roles, 37.5 to 60% irreversible harm ungated versus 2.5% with an enforced gate, and 94% of leaders confident versus 32.7% enforcing least privilege.
Part 3 · 16 October 2026

Agentic architecture patterns, part 3: the control plane, the patterns that decide whether an agent may run

A rule the agent reads is a request: 93% of permission prompts approved, 13.6% refused a planted command, 0.03 ms for a policy decision

A control is something the agent cannot read. Nine patterns that sit outside every agent, with what each costs in milliseconds, the measurements, and the incident of 2025 or 2026 each would have caught.

56 min read
Infographic of the Nightly harness as a graph around a loop: a 2 a.m. routine and a workflow script running 8 upgrades at once feed a Stop-hooked loop capped at 30 turns and $2 a pass, then a test gate, a durable approval interrupt and a pull request as the only write, with a $80 worst night. Below, a log ruler of the caps platforms ship, from 3 sub-agent levels and 8 Stop-hook blocks to 1,000 agents per workflow run and 4,096 items per parallel call, with the three settings that default to no limit.
Part 4 · 17 October 2026

Agentic architecture patterns, part 4: harness engineering, hands-on, with the loops and graphs that run the model

Eight Stop-hook blocks, 1,000 supersteps, 1,000 agents a run, a seven-day fuse, and two caps that default to none

The harness is now the product. A hands-on guide to loop engineering and graph engineering: what stops a loop, what a graph saves, and a complete harness built piece by piece with the numbers that bound it.

49 min read

IIRecursive self-improvement

A three-part series on machines that improve the process that improves them: the theory from Good’s intelligence explosion to the Gödel machine, the self-play and self-training loops that work and the verifier that bounds them, and code that rewrites its own coder, with the evidence on what goes wrong and what to build.

IIIThe summer the agents got out

Between July and September, three labs’ agents reached systems they were never meant to touch. What failed, what the independent investigators found, and what a platform team should change before it runs agents of its own.

IVPlaybooks for the platform team

Reference designs written to be built from: a router that turns a price change into a config change, a bench that decides which model runs each task, identity for things that never log off, grounding as the product, classification before everything, and how to onboard an agent like a hire.

8 essays
Six task classes of a procurement email triage placed on a latency-budget axis from 2 seconds to 24 hours, as bubbles sized by cost per 1,000 runs (about $2, $84, $14, $14 and $1, plus a local step with no token bill), against time-to-first-token lines of 0.9, 16.5, 44 and 51 seconds; with the routed total of about $115 against $170 on one flagship, a ten-week timeline of six rate cards, and cache, batch and tokenizer numbers.
Platform design · 25 June 2026

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

Six rate cards in ten weeks; $115 against $170 per 1,000 runs

Six frontier rate cards in ten weeks, one suspension and one no-show. The fix is a router that routes on a declared task class, with the registry, cache discipline and cost ledger to match.

19 min read
A slope chart of the invoice-triage worked example: Gemini 3.5 Flash costs 3.2 cents per task against 8.4 for GPT-5.5 and 7.4 for Opus 4.8, but with escalations priced at $4 its all-in cost is 29 cents against 24 and 21, shown with token counts, success and escalation rates, the July 2026 rate card, and a panel on judge reliability (position bias of 0.192, 0.125 and 0.002, kappa 0.511 against 0.849 exact match).
Evaluation · 3 July 2026

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

Exact-match agreement overstates judge validity by 33 to 41 points

Model judges agree with themselves at 0.99 and with the truth far less than exact match implies. The fix is not a better leaderboard but a bench: sampled tasks, audited graders, a rate card, three numbers.

19 min read
A badge for the worked example, the finance reconciliation agent (controller as sponsor, three access packages, one-hour and five-minute credentials, ninety-day review, retired a year later), beside a matrix of nine onboarding steps against Entra Agent ID, AgentCore Identity, SPIFFE and SPIRE, the Governance Toolkit and Zenphi or IBM showing which product helps at each step and what each leaves to the platform team, above key numbers such as Gartner's 40 percent and OWASP ranking identity third of ten risks.
Security · 8 July 2026

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

Entra demands a sponsor at creation; nobody else does

Couchbase, IBM and Zenphi shipped governed agents in one week; none shipped identity. A prompt is not a principal. Here is the joiner, mover, leaver discipline for agents, with a finance reconciliation agent onboarded end to end.

20 min read
A six-stage enterprise retrieval pipeline with GaRAGe's number at each stage: 65 million downloads for the query gate classifier, 47.8 percent relevant passages on Enron against 85.6 on the Web, a best relevance-aware factuality of 60.67, a best attribution F1 of 58.9 and a best deflection rate of 31.1 percent, with the passage label mix and where stale, sparse and over-summarised failures surface.
Retrieval · 16 July 2026

Grounding is the product: what GaRAGe shows about enterprise RAG

The best model grounds 60% of answers and deflects on 31%

Amazon’s GaRAGe benchmark found the best model grounds 60 percent of answers in relevant passages and deflects on 31 percent of the questions it should. A footnote cites my classifier. What both mean for enterprise RAG.

18 min read
Data infographic: one disputed-transaction email traced through an agent across four class bands, rising from confidential to regulated when the statement attachment is classified, dropping to internal through a named declassification for triage, rising again to regulated when core-banking rows return, and splitting memory into regulated and internal partitions, with each class's tier, residency and retention rules, the five dated rules that presume a class, and the 1,200 agents and 141,006 runs from the summer's incidents.
Platform design · 24 August 2026

Classify first: the control every other AI control depends on

A prompt inherits the highest class of anything in it

Routing, residency, de-identification, retention and agent permissions all run on a data class, and when the class is missing they run on a default. A scheme, a classifier and a worked example for building classification first.

19 min read
Infographic: an onboarding badge for an always-on procurement agent (owner, 90-day review, access, alone/approval/never rules, budget alert at 80 percent, 50 actions sampled weekly, stop under 60 seconds) beside a matrix of identity, boundary, budget and kill switch against what Microsoft, NVIDIA and OpenAI shipped and the gaps left, a to-scale bar of 164 minutes to stop by hand against the 60-second target, and key numbers: Argon's 1,000,000-token output, 4,000+ Dots apps, GPT-6.1 Sol at $2 and $10, and five incident-review questions.
Agentic systems · 2 October 2026

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

Four controls an always-on agent needs; no vendor ships all four

Dots, the Agents API, Gemini 4 Argon and NVIDIA’s hardware-enforced safety platform landed in one week. Before you run an always-on agent, give it what you give a new hire.

14 min read
Infographic: documents, images, audio and video, each with its benchmark result (96.96 on OmniDocBench, 68.13 on photographed pages, 95.02 on charts, 78.08 on audio, 50.70 on text-rich video timing), flow through the 5B Youtu-Parsing-Omni model into one annotated JSON record whose fields note the box grid, second-level timestamps, unknown languages and the missing confidence field. Below, Gemini 3 Pro's OmniParsingBench scores moving by up to 30 points between two scorers, and adapters falling from 15 to 8 under one contract.
Enterprise AI · 12 October 2026

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

One JSON record for every input. Change the scorer and Gemini 3 Pro’s video score moves 30 points

A 5B open model turns pages, charts, calls and videos into one grounded JSON record. The schema is the advance; the scores swing by 30 points with the scorer, and confidence is missing.

24 min read
Infographic: on the left, BrowseComp accuracy climbing from 51.5 percent in April 2025 to 90.4 percent for a single agent in July 2026, above a human baseline of 29.2; on the right, the share of GPT-5.4's cited facts that the page supports falling from 78.6 percent at 2 tool calls to 16.7 at 150, while links keep resolving. Beside them, four models whose links resolve 94 to 100 percent of the time but whose pages support the claim 47.7 to 76.8 percent, and three key numbers on fabricated URLs, judge agreement and report quality.
Evaluation · 15 October 2026

How to evaluate a deep research agent, and why the eval deserves more time than the agent

BrowseComp 51.5% to 90.4% in fifteen months; cited pages support the attributed fact 47.7% of the time, 17% at 150 tool calls

BrowseComp is near saturation at 90.4 percent, yet cited pages support the attributed fact as little as 47.7 percent of the time. Evaluate retrieval, answers, claim support and deflection separately, grade the grader, and design the eval before the agent.

58 min read

VRunning it

The operational side nobody writes the vendor deck about: what on-call looks like when the pager is for a model, who pays the AI bill, how to read a model card like a contract, what happens when the frontier model you built on is switched off, and whether the open model you downloaded is carrying a trigger.

5 essays
Data infographic: a 12-hour on-call shift drawn as a clock dial with three tier rings, marking one hypothetical night for an invoice agent (a breaker at 23:52 with no page, a 14.4-times cost burn paged at 01:15 and paused in 15 minutes, a blocked egress at 03:10 stopped in 48 seconds) inside a pager budget of two incidents per shift. Panels list four objectives (97 percent, 2 to 6 percent, under 4 minutes, 1.3 times cost), OpenAI's July record from an alert waved through on 27 June to a first stop 5.5 hours after the 19 July alert and a last shutdown 10 days later, burn-rate alert thresholds and the 72-hour to 15-day reporting clocks.
Operations · 31 August 2026

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

A monitor outside the agent would have paged a day earlier

When the pager is for a model, the failures are not outages. The signals, SLOs, pager tiers, kill switch, incident clock and postmortem template for an on-call engineer, drawn from OpenAI’s 26 August report, Microsoft’s toolkit and the NCSC.

20 min read
A treemap of one illustrative month's AI bill of about $19,000 split by owner: customer care $10,009 at 1.7 cents a conversation, consumer lending $5,362 at 4.5 cents an application, procurement $1,752 at 2.9 cents an invoice and an unowned platform line of $1,878, with the Luna line at $6,960 on 9 July prices against $1,392 on 17 August prices, a timeline of six price events in six weeks and FinOps adoption rising from 31% to 98%.
Unit economics · 17 August 2026

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

Tokens, not hours: a unit finance has never billed before

Luna was cut 80 percent three weeks after launch and Kimi K3 prices a cache hit at a tenth of a miss. An AI bill needs a tenant on every call and a ledger priced the day it happens.

19 min read
Data infographic: the twelve fields to extract from a model release as a contract schedule, each with where it lives, what was copied from the Claude Fable 5.1 card (five prices from $50 output to a $0.25 cache read, a tokenizer making about 30 percent more tokens, over-refusal of 0 and 0.34 percent, Terminal-Bench 56 percent at max effort, retirement not before 1 September 2027), what TypeSafe AI's Jev page gives ($0.042 per million input, 67.8 percent agreement) and does not (cutoff, deprecation policy, region, architecture), and the registry, bench, calendar or held-out test each field lands in.
Model strategy · 17 September 2026

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

212 pages of system card, twelve fields worth copying

The Fable 5.1 system card runs to 212 pages and the price is not in it. Twelve fields to extract from any release, benchmark settings to read as a sceptic, and what to demand when a card withholds its architecture.

20 min read
A five-step chain of how an open-model backdoor is planted and fired (4.47M hub versions scanned, 125 of 625 fine-tuning rows poisoned for about $8 and 0.6% of the weights, 50/50 clean prompts correct, 50/50 triggers fired, four exfiltration steps), with the control at each step and what it catches or misses, and a ruler of poisoned samples needed: 10, about 100, 125 and 250.
Security · 10 October 2026

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

Under $50, 100% trigger rate, 100% clean accuracy: the benchmark reports a healthy model

ProjectDiscovery backdoored an open model for under $50. It passed every benchmark and leaked credentials on a hidden phrase. Why scanners miss it, and the controls that contain it.

22 min read
A descending staircase of notice on a log scale, from 6 to 12 months on a partner platform through 60 days on Anthropic's API (62 and 61 kept) and 14 days for a rationing change to none for the 12 June export-control suspension, beside a timeline of Claude Fable 5's three days on sale and a split comparing the Friday loan-review batch without a fallback chain (8,000 documents queued) and with one ($381 a night on Opus 4.8 against $762).
Model strategy · 13 June 2026

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

Switched off for every customer, three days after launch

Fable 5 launched on Tuesday and was switched off for the world on Friday by a US export-control directive. The three warnings that came first, what broke, and the architecture that kept running.

19 min read

VIModels and money

What the labs shipped and what it costs to use: token efficiency as the new benchmark, a price war that halved the frontier in a week, an Opus at the old price, open weights you can hold, capability gated by who you are, a new class of decision models, and a phone that runs twenty billion parameters.

9 essays
Squares sized by cost show one 40-step agent task falling from $0.45 in a naive loop to $0.15 with prompt caching, $0.13 with 17% fewer output tokens, $0.08 with 25 steps instead of 40 and $0.04 with Flash-Lite on the 60% of bounded steps, beside an invoice agent's monthly bill shrinking from $27,000 to $2,500, above a strip of July's releases and prices from GPT-5.6, Inkling, Kimi K3, Gemini 3.6 Flash and Flash-Lite.
Unit economics · 22 July 2026

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

17% fewer output tokens is now a launch headline

The labs stopped competing only on intelligence in July and started competing on what a task costs. Tokens to done is the benchmark that matters.

14 min read
Data infographic: input prices per million tokens on a log ruler, from GPT-6 Luna at $0.10 (cached $0.01) to Fable 5.1 and GPT-6 Astra at $10, with Opus 5.5 cut from $5 to $4 and its cache reads 60 percent to $0.20, above an illustrative monthly bill for 200,000 supplier cases falling from $6,800 on Opus 5 to $5,300, $1,050 when routed by task class and $880 with caching. Side panels show three dated price changes and Sonnet 5.5 scoring within two points of Opus 5.5 at the same $2 price.
Unit economics · 29 September 2026

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

Cheapest frontier model is 100× cheaper than the dearest on input

Opus 5.5, GPT-6 Sol and Luna, Sonnet 5.5: the frontier’s prices halved in a week. A price war only helps if your architecture can route to the cheaper tier.

15 min read
Four stacked columns show one bug-fix class at an unchanged $5 and $25 rate card costing $2.77 per completed task on Opus 4.8 at xhigh, $2.00 on Opus 5 at high with the prompt cleaned, $2.65 on Opus 5 at xhigh and $3.19 on Opus 5 with the old prompt, each split into fresh input, cache reads, output and failed tries at 74 to 84 percent pass rates, beside Anthropic-reported score gains on SWE-bench Pro, OSWorld 2.0, Frontier-Bench and ARC-AGI-3.
Model strategy · 25 July 2026

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

96.0 on SWE-bench Verified at the same $5 and $25

Opus 5 arrived at $5 and $25, the same list price as Opus 4.8, and beat Fable 5 on several agentic benchmarks. Three variables moved anyway. Here is the release-day protocol that finds them.

20 min read
A break-even chart of monthly cost against utilisation of a 64-GPU B200 cluster: Kimi K3's API at $0.85 per million and Claude Opus 5's at $1.54 rise as slopes against flat self-hosted lines at $135k reserved and $287k on demand, crossing at 22, 40, 46 and 84 percent, with the 200-engineer workload at $38,300 on Kimi and $69,300 on Opus 5, beside a memory ruler showing the 1.56 TB checkpoint against one 1.4 TB node, the cache prices and the licence thresholds.
Open weights · 28 July 2026

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

2.8 trillion parameters, 104 billion active, $0.30 per million cached

Moonshot published Kimi K3’s weights on 27 July under a licence that costs internal users nothing. Holding 1.56 terabytes is cheap; running them is a sum most teams have not done, shown here with the licence read clause by clause.

21 min read
Data infographic: the week of 1 to 3 September 2026 as three day columns split by a gate, with what anyone can buy above it (Claude Fable 5.1 at $10 and $50 with $0.25 cache reads, Gemini 3.8 Flash at $0.75 and $3.75 until it doubles on 1 January, GPT-6 Astra at $10 and $50, Muse Spark near $0.10) and what needs vetting below it (Mythos 5.1, Gemini 3.8 Flash Cyber with 650-plus partners, Astra's Critical-rated cyber tier). A side panel shows a 119-fold price spread, 85 and 60 percent fewer refusals and interventions, and a buyer's calendar of price dates.
Model strategy · 4 September 2026

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

Four frontier models in three days; the top tiers now gated by who you are

Anthropic, Google, OpenAI and Meta shipped in 72 hours. The real change is that the most capable tiers are now gated by trust, not price.

17 min read
A log-scale ruler of decision latency from 28 ms (CLM-8B) and 39 ms (Clef-flash) through Jev's 0.4 s to 37.8 s for Claude Opus 5, a staircase of thirteen System One model releases between 15 September and 1 October 2026, and TypeSafe's comparison of Jev with GPT-5.6 Terra and Claude Opus 5 on accuracy (67.8%, 67.9%, 73.1%), cost per case and latency, with three findings from the independent study.
Enterprise AI · 5 October 2026

The fast half of intelligence is getting its own models

Twelve decision models launched in six weeks

Three years of making models think slower produced real gains and a real bill. System One models, built to make bounded decisions in a single pass with calibrated confidence, are the correction.

18 min read
A bubble chart of the five decision-model fine-tunes in Unsloth's guide, plotting test accuracy (77% to 81%) against training time (10 to 49 minutes) with bubble area for peak GPU memory (2.5 to 14.4 GB), beside before-and-after slopes on three test sets (typed decisions 36% to 73% and 33% to 78%), the 0.6 and 0.85 probability thresholds of the worked example, Laya's typed-decisions scores and the $46,000 a year Jev list-price comparison.
Enterprise AI · 8 October 2026

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

A 0.8-billion-parameter model went from 36% to 73% in 42 minutes on 4 GB

Unsloth now trains decision models from any language model and serves Laya locally at a Jev-compatible endpoint. The forward pass, the loss, the calibration step and a ticket-routing fine-tune, read from the source.

26 min read
A phone holding a grid of 200 dots, each 100 million parameters of Apple's 20-billion-parameter AFM 3 Core Advanced, with 10 to 40 lit for the one to four billion active per request, beside the three inference tiers (device with a 4,000-token context, private cloud with 32,000, frontier API) and what may cross the two boundaries between them, and key numbers: a 12 GB memory floor, 15 supported languages with no Indian one, Apple's 9 and 28 percent of India's phone market, and Gemma 4 12B at 11.95 billion parameters.
Model strategy · 10 June 2026

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

20 billion parameters on a phone, 1 to 4 billion active

A 20-billion-parameter model on the iPhone, Private Cloud Compute on Google Cloud, Gemma 4 12B on a laptop. The device is now an inference tier, and the design problem is routing across three of them.

19 min read
Infographic: a staircase of measured single-user speeds for very large models, grouped by where the weights sit, from AirLLM streaming Kimi K3 from disk at 292 seconds a token, through experts streamed from SSD at 1.83 to 4.36 tokens a second and experts in system memory at 9.2 to about 30, to DeepSeek-V4-Flash at 39.4 tokens a second in a laptop's unified memory, a gap of over 10,000 times. Beside it, the bandwidth ladder from a 7 GB/s SSD to a 3,350 GB/s H100 with the ceilings it sets for dense and mixture-of-experts models, and the time to label 2,000 documents: 19 hours to 4 weeks with AirLLM against 45 minutes to 3 hours on a 128 GB machine.
Open weights · 14 October 2026

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

Kimi K3 in under 4 GB at 292 seconds a token; a 284B model on a laptop with 128 GB of RAM at 39 tokens a second

Running a giant open model on a small machine is two problems. Fitting is solved. Speed is memory bandwidth divided by bytes per token, which is why MoE models in fast memory win and AirLLM suits only batch jobs.

38 min read

VIIAgents at work

Agents that drive a screen, agents that answer the phone, and the controls that let either one near a customer.

2 essays

VIIIRules and regulators

The duties that arrived on time, the ones that were deferred, the one control plane that satisfies three regulators at once, and what India’s DPDP Rules and RBI’s FREE-AI framework ask of an AI platform.

4 essays
Four product mock-ups pin the EU AI Act's Article 50 duties where they live (a first-turn AI disclosure, an emotion-analysis notice, a C2PA mark and visible label on a campaign image, a named-editor gate on a published report), beside SynthID-Text detection F1 falling from 1.0 to 0.884, 0.842, 0.788 and 0.711 under attack, a timeline from the 17 December 2025 draft to the 2 August 2026 start and 2 December runway, and the Article 99 penalty of €15M or 3% of turnover.
Governance · 3 August 2026

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

Fines up to EUR 15 million or 3% of turnover, from 2 August

Article 50 of the EU AI Act arrived on 2 August while the high-risk chapters were deferred. Disclosure, marking and labelling live in the product, not the compliance file. What to build once for Europe and India.

19 min read
A matrix maps the NCSC's seven agent-safety considerations to the EU AI Act articles and RBI FREE-AI and DPDP duties each satisfies, emergency shutdown highlighted in clay, beside one incident clock (2 days, 72 hours, 10 days, 15 days), one 12-month retention setting against the 6-month and one-year floors, and the RBI survey (about a third board oversight, 18% audit logs, 14% real-time monitoring).
Governance · 20 August 2026

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

Seven NCSC considerations map onto one control plane

The NCSC, the EU AI Act and the RBI converge on the same engineering controls for agents. A platform team should build them once as a control plane and let each regulator read the evidence.

19 min read
Infographic: all 26 FREE-AI recommendations as tiles under their six pillars, with the thirteen that name regulated entities in clay (Rec 10 and every one from 14 to 25) and the build list's clocks marked (annual policy review, semi-annual red teaming, 1% human sample, DAKSH in 6 hours, half-yearly inventory, audit review every 2 years), a rising staircase from the 13 August 2025 report to the 24 June 2026 draft guidance, 31 July cyber Directions and 11 August FIBAC speech, and the RBI survey: 20.80% of 612 entities using AI, a third with board oversight, a quarter with incident processes, 85% wanting a framework.
Governance · 1 October 2026

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

Regulated entities are named as an owner in 13 of FREE-AI’s 26 recommendations

The RBI’s FREE-AI report set seven sutras and 26 recommendations for AI in finance. All 26 by pillar and owner, what the RBI has done since, and a ten-item build list for banks and NBFCs.

21 min read
A month-by-month calendar strip from November 2025 to May 2027 marking every dated DPDP step (notification on 13 November 2025, the January 2026 consultation, Rule 4 on 13 November 2026, 221 days from 4 October 2026 to 13 May 2027), above seven tiles mapping where personal data sits in an LLM support agent to the control and rule each needs, maximum penalties of ₹250, 200, 200 and 150 crore, and the 72-hour, one-year, 90-day and three-year clocks.
Governance · 4 October 2026

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

Most DPDP Rules obligations reach AI systems on 13 May 2027, eighteen months after notification

India’s DPDP Rules never mention AI, yet from 13 May 2027 they govern every LLM system that touches personal data. A map of consent, erasure, logs, breach reporting and SDF duties onto prompts, indexes and vendors.

19 min read

IXIndia and sovereignty

The tokeniser tax on Indian languages, open weights at trillion scale, and the week residency became a feature the hosted labs sell. What sovereignty costs, and which layer of it you are actually buying.

5 essays
A dumbbell chart of tokens per word for English and ten Indian languages on the GPT-4o tokeniser against Krutrim's Indic tokeniser, from 1.33 against 1.48 for English to 6.39 against 1.80 for Odia, beside reply-cost multiples (about 1.2 times for Hindi, 2.4 for Tamil, 4.8 for Odia), words that fit in a 128,000-token window (96,000 English, 20,000 Odia), the token counts of one support turn, and the IndiaAI Mission's compute and funding numbers.
Sovereign AI · 16 June 2026

Indic models: the gap between the benchmark and the bazaar

An Odia reply costs 4.8× an English one on a frontier tokeniser

An Odia reply costs a frontier API almost five times an English one before the model has done anything. India’s sovereign models fix that, but no two report the same benchmark. What a buyer should measure instead.

19 min read
Nested squares sized by total parameters compare Kimi K3 (2.8T, 1.4 TB, 64+ accelerators), Qwen3.8-Max (2.4T, 86.6 against 88.8 on TerminalBench 2.1, no licence), Inkling (975B, Apache 2.0) and a tiny Nemotron 3.5 Lightning (31.6B, 25 GB on one GPU), beside a five-layer sovereignty table in which open weights supply weights and data but not always licence, and neither compute nor controls, and India's IndiaAI figures of 93 lakh GPU-hours and 20 proposals with five released.
Sovereign AI · 14 August 2026

Open weights at trillion scale, and what sovereignty actually costs

Five layers of sovereignty; open weights give you one

Four open-weight models at or near frontier scale in five weeks. What they give an enterprise, what they do not, and how to decide per workload.

14 min read
A two-lane timeline from 27 August to 5 October 2026 showing which models reached Bedrock's India profile (GPT-5.6 Terra and Luna, then Claude Opus 5, Sonnet 5 and Haiku 4.5) while Fable 5.1, Opus 5.5, GPT-6 and Sonnet 5.5 went global only, with an illustrative bank's monthly costs in India against offshore ($13,500 against $8,500 for chat, $430,000 against $170,000 for a coding agent), the five layers the India profile does and does not cover, and the price gaps behind them.
Sovereign AI · 6 October 2026

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

Residency costs a model generation: Opus 5 in Mumbai, Opus 5.5 everywhere else

Claude Opus 5, Sonnet 5 and Haiku 4.5 now run in-country in India on Amazon Bedrock. What that buys a regulated enterprise, what it costs in model generation and price, and the five layers residency does not cover.

21 min read
Data infographic: a 30-day evaluation calendar in four gated weeks, from building a 2,200-item test set (400 Hindi, 500 Hinglish, 400 Tamil, 400 Tanglish, 200 English, 300 calls) through automatic gates (98 percent script match, 90 percent accuracy), native-speaker rating and safety tests to a 5 percent pilot decided on cost per resolved conversation. Beneath it, the evidence: Indian languages at 2.1, 8 and 13 times English token counts, judge agreement with native raters falling from 0.54 to 0.24, and a judge that picked a winner 87 percent of the time when both answers were hallucinated.
Evaluation · 26 September 2026

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

In ten Indian languages, a model judge agreed with native speakers at a kappa of 0.31 on direct scores

Before you choose a model for Indian-language users, test it on your own traffic. A 30-day plan: romanised and code-mixed test sets, tokens per word, native raters, judge calibration, safety in every script and a written decision rule.

19 min read
A matrix of four residency options against five hops of one request (gateway, inference, logs, safety copy, add-ons) showing which stay in India, beside chips listing the models each cloud processes in India on 7 October 2026 and the newer ones it does not, and six figures from the rules and terms: RBI's 24 hours, CERT-In's 180 days of logs, 30- and 90-day safety copies, DPDP Rule 13(4) from May 2027 and 38,000 IndiaAI GPUs.
Sovereign AI · 7 October 2026

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

From Mumbai, Bedrock’s APAC profile can route to Tokyo, Seoul, Osaka, Singapore and Sydney. APAC is not India

Which models run with processing inside India on Amazon Bedrock, Azure OpenAI, Vertex AI and Indian GPU clouds in October 2026, what DPDP, RBI and SEBI actually require, and where residency claims break.

20 min read

XResearch and field notes

A frontier model that patched another model’s weights by hand, a Millennium Prize proof nobody can read, and ten years of teaching machines to talk, from banking chatbots to foundation models.

4 essays
Infographic of the MIT CSAIL weight-space programming result: a strip of Qwen3.5-0.8B’s 24 decoder layers with layers 15 and 19 marked as the two edited by GPT-6 Astra, the share of weights changed (0.016%, 2 layers × 64 columns × 1,024 rows = 131,072), before-and-after scores from the 50-pass edit (MATH-500 7.9% to 48.1%, GSM8K 13.8% to 67.1%, IFEval 21.0% to 62.3%, answers hitting the token limit 93.0% to 33.1%), the 10- and 25-pass runs that barely moved, the Gemma 3 270M control that stayed between 6.8% and 7.7%, and a ledger of all eight experiments with their results, from identifying games from 17,253 weights to writing a 9×9 Go policy from zeros.
Research · 3 November 2026

Weight-space programming: what a 131,072-weight patch proves, and what it does not

131,072 of 800 million weights, two layers, no gradient: MATH-500 7.9% to 48.1%, and 93% to 33% of answers cut off

A frontier model read a small model’s weights, found the loop that stopped it finishing, and patched 0.016% of them by hand. What the eight experiments show, what the Gemma control proves, what it cost, and why provenance just became a control.

47 min read
Layer-by-layer strips of nine hybrid models from Jamba (4 of 32 layers attention) to Kimi K3 (24 of 93) and Qwen3.8-Max (23 of 92), showing every one kept some attention, beside the author's memory example for a one-million-token sequence (256 GB all attention, 64 GB for a 3:1 hybrid, 64 MB all linear), the Kimi Linear ratio ablation and four key numbers including the 7B largest pure Mamba.
Research · 8 October 2026

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

Linear time, constant memory, and still not the architecture of a single frontier model

Three years after Mamba, no pure state space model runs at the frontier, yet Kimi K3, Qwen3.8-Max and Nemotron 3.5 Lightning are three-quarters linear layers. The papers behind the paradox, and a million-token memory sum.

29 min read
Data infographic: the Navier-Stokes result as five gates, the machine ones cleared (a run of about 10,000 agents over 88 hours with 2.7 million messages and 130 billion tokens, about $6.5 million at list price, then a 17-hour Lean check) and the human ones not (a 166-page proof called incomprehensible, a forced-versus-unforced scope and credit dispute, and a Clay Institute two-year rule). A log ruler of hours below sets 17 hours in Lean against 40 person-hours for one page by hand and about 133,000 for all 166 pages.
Research · 12 September 2026

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

10,000 agents, 88 hours, one proof nobody can read

OpenAI’s Navier–Stokes claim is a preview of research at swarm scale. The proof was the easy part; verification, credit and understanding are the hard ones.

16 min read
A ten-year timeline of the author's roles from Active.Ai to Tata Group with each era's numbers (95 percent accuracy on enquiries, 23 percent of 15 billion Haptik conversations not in English, a 35-person Salesken team, Indic text at about 1 percent of Common Crawl) and the stack layers above it, over a ledger of seven control points comparing the intent era with a typical June 2026 agent stack, from per-intent thresholds such as card_block at 0.94 to per-token prices of $1.50 and $9 for Gemini 3.5 Flash and $5 and $25 for Opus 4.8.
Conversational AI · 4 June 2026

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

95% accuracy on bounded enquiries across 300 million interactions

From intent classifiers for banks to LLM agents: ten years of building talking machines, and the five controls the intent era enforced that the agent stacks of June 2026 have quietly dropped.

19 min read

Publishing calendar

48 essays since June 2026
43 publishing days · an essay in each of the last 1 weeks

The archive

Every essay, newest first
  1. 3 November 2026Weight-space programming: what a 131,072-weight patch proves, and what it does notResearch47 min
  2. 17 October 2026Agentic architecture patterns, part 4: harness engineering, hands-on, with the loops and graphs that run the modelAgents49 min
  3. 16 October 2026Agentic architecture patterns, part 3: the control plane, the patterns that decide whether an agent may runAgents56 min
  4. 15 October 2026How to evaluate a deep research agent, and why the eval deserves more time than the agentEvaluation58 min
  5. 15 October 2026Agentic architecture patterns, part 2: many agents, and the ways they fail togetherAgents62 min
  6. 14 October 2026Agentic architecture patterns, part 1: one agent, one loop, and the patterns inside itAgents67 min
  7. 14 October 2026Big models on small machines: fitting is solved, speed is a bandwidth billOpen weights38 min
  8. 12 October 2026Youtu-Parsing-Omni and the perception contract: adopt the interface, not the assumptionEnterprise AI24 min
  9. 11 October 2026Recursive self-improvement, part 1: the idea, the Gödel machine, and the wall it hitResearch23 min
  10. 11 October 2026Recursive self-improvement, part 2: self-play, self-training, and the verifier ceilingResearch23 min
  11. 11 October 2026Recursive self-improvement, part 3: code that rewrites its coder, from FunSearch to AIDE²Research25 min
  12. 10 October 2026A $50 backdoor in an open model: the mechanism, and the defenceSecurity22 min
  13. 8 October 2026The architecture that won the argument: state space models, three years onResearch29 min
  14. 8 October 2026Train your own decision model: the decision tier is now a 40-minute fine-tuneEnterprise AI26 min
  15. 7 October 2026LLM data residency in India: what runs in-country in October 2026, and where it breaksSovereign AI20 min
  16. 6 October 2026Claude answers from Mumbai: residency is a product now, and it is not sovereigntySovereign AI21 min
  17. 5 October 2026The fast half of intelligence is getting its own modelsEnterprise AI18 min
  18. 4 October 2026DPDP Rules for AI systems: what the 18-month clock asks of your models, prompts and logsGovernance19 min
  19. 2 October 2026An always-on agent is a hire. Treat it like one.Agentic systems14 min
  20. 1 October 2026RBI’s FREE-AI framework, explained: seven sutras, 26 recommendations, one build listGovernance21 min
  21. 29 September 2026The price of intelligence halved in a week. Your architecture should notice.Unit economics15 min
  22. 26 September 2026Indic LLM evaluation: a 30-day test plan for Hindi, Tamil and Hinglish usersEvaluation19 min
  23. 23 September 2026The agents got out. Here is the timeline, and what a platform team should change.AI safety in production16 min
  24. 17 September 2026Read the card like a contract: twelve fields from the system card to the registryModel strategy20 min
  25. 12 September 2026Ten thousand agents, 88 hours, and a proof nobody can readResearch16 min
  26. 4 September 2026Four frontier models in three days, and the one change that mattersModel strategy17 min
  27. 31 August 2026On call for a model: the runbook when the pager is for an agentOperations20 min
  28. 27 August 2026Cheating on the test: the Hugging Face agents were reward hacking firstEvaluation integrity18 min
  29. 24 August 2026Classify first: the control every other AI control depends onPlatform design19 min
  30. 20 August 2026Two regulators, seven controls: build the agent control plane onceGovernance19 min
  31. 17 August 2026The AI bill nobody owns: a chargeback playbook from the platform seatUnit economics19 min
  32. 14 August 2026Open weights at trillion scale, and what sovereignty actually costsSovereign AI14 min
  33. 10 August 2026Voice is the bottleneck: no language model recovers a wrong product nameConversational AI19 min
  34. 7 August 2026Three labs, one vendor: the sandbox that wasn’tSecurity14 min
  35. 3 August 2026The label that did arrive: the one AI Act duty that came on time lives in the productGovernance19 min
  36. 30 July 2026An AI lab’s own models broke into Hugging Face. Read the postmortem.Security13 min
  37. 28 July 2026Kimi K3: the weights you can hold, and the arithmetic before you doOpen weights21 min
  38. 25 July 2026Opus 5 at the old price: an unchanged rate card is not an unchanged billModel strategy20 min
  39. 22 July 2026Tokens to done: the benchmark the labs started competing on in JulyUnit economics14 min
  40. 16 July 2026Grounding is the product: what GaRAGe shows about enterprise RAGRetrieval18 min
  41. 8 July 2026An agent is a principal: identity is the control no vendor ships across vendorsSecurity20 min
  42. 3 July 2026The evaluation bench, in full: how to decide which model runs each taskEvaluation19 min
  43. 27 June 2026The browser is the API now, and RPA’s risks are back at ten times the scaleAgentic systems19 min
  44. 25 June 2026Routing by task class: a reference design for the enterprise model routerPlatform design19 min
  45. 16 June 2026Indic models: the gap between the benchmark and the bazaarSovereign AI19 min
  46. 13 June 2026The model that switched off: a frontier model is now a dependency on a day’s noticeModel strategy19 min
  47. 10 June 2026Twenty billion parameters in your pocket: the device is now a tierModel strategy19 min
  48. 4 June 2026Ten years of talking machines: what the intent era got rightConversational AI19 min