Ivan Nosov

AI is reshaping knowledge work. I want the people who do the work to shape it.

I'm Ivan Nosov. I lead AI transformation: I set the direction, design the system around the work, and turn the people who do the work into the ones who build it. Here I write down how it's done.

Method
What to write down, who builds, who owns it, and what to fund.
Practice
Notes from building it, including what failed.
Evidence
179 graded sources behind the claims, and a new piece every week.

Read the latest piece Write to me on LinkedIn This week in AI

Ivan Nosov explaining with both hands held level, glasses, white shirt, FAIR lanyard.
Speaking at FAIR Barcelona, 1 October 2026. Photo: FAIR Barcelona.

Latest writing

All writing

My newest pieces, numbered and dated: what I make of what happened, with the source it starts from and a plate from the record beside the lead.

No. 62·

The tasks worked. The job did not.

Epoch AI tested six models on its own work. Where they failed is where most companies run on unwritten rules.

Read No. 62

Epoch AI report page titled Can AI automate Epoch?, dated 8 October 2026, listing three key findings. The second, on models missing implicit conventions such as visual style and audience interests, is highlighted.
Exhibit Can AI automate Epoch? Key findings section; highlight added by me. Source

  1. No. 61

    OpenAI's 722 maths papers and the shape of jobs

    This week mathematics showed how the shape of many jobs will change, and it happened faster than anyone planned for.

    Business Insider article headline about an NYU professor saying OpenAI's release of maths papers wiped out early-career researchers' projects, with the byline and a quote from Tristan Buckmaster highlighted.
    Exhibit Business Insider, 10 October 2026. Source

  2. No. 60

    The Italian Hackathon League final at OGR Torino

    An unusual start to the day: the final of the Italian Hackathon League, at OGR Torino during Wave.

  3. No. 59

    FAIR Barcelona 2026: keynote recap

    Last Thursday I opened my speech at FAIR Barcelona with a question I have been solving for a while: what would it take to build JARVIS for HR?

  4. No. 58

    Everyone becomes a bit of an engineer

    Almost everyone who gets serious about AI eventually becomes a bit of an engineer.

    Editorial illustration: a wrench cut out of a printed page on an orange ground, headline AI makes you a bit of an engineer.
    Exhibit Visual from my LinkedIn post. Source

  5. No. 57

    Twenty supplier offers: where the change AI brings is hard to reach

    Fundamental change that AI brings might sound simple, yet it is hard to reach.

  6. No. 56

    The more I delegate to AI, the more distilled knowledge is worth

    The more work I delegate to AI, the more valuable the knowledge I have taken the time to distil becomes.

    One-page diagram titled Graph engineering: a process drawn as a network of nodes from frame and evidence check to storyboard, integration and release, with distilled knowledge feeding every node.
    Exhibit Visual from my LinkedIn post. Source

  7. No. 55

    Databricks as an AI workbench for knowledge work

    Databricks became an end-to-end AI workbench for knowledge work. More than 100 knowledge workers now use coding-class agents on it against challenges of HR, Finance, Commercial, and Marketing.

Top reads in the Library·Long reads·All 62 pieces

Numbers worth knowing

All 481 figures on Numbers

Figures from the highest-graded sources, with the chart behind each one. Every number links to its source.

89%

of firms report no AI effect on labour productivity over the past three years, while 69% use AI.

Firms using AI, by country
  • US 78%
  • UK 71%
  • Germany 65%
  • Australia 59%

S NBER,

19%

employment shortfall for 22-25-year-olds in AI-exposed occupations against less-exposed peers, June 2026.

Employment change, ages 22-25, Nov 2022 to Jun 2026 (approximate, as the source gives it)
  • Two most exposed quintiles -11%
  • Three least exposed quintiles +10%

S Stanford DEL,

80%

report measurable impact where AI strategy is clear and tools are limited; 60% where tools are strong and direction is not.

Measurable impact, by condition
  • Limited clarity, limited tools 55%
  • Limited clarity, strong tools 60%
  • Strong clarity, limited tools 80%
  • Strong clarity, strong tools 83%

S BCG,

93%

of workers have used chatbots where employers combine encouragement, tools and training; 41% where employers do nothing.

Workers who have used chatbots
  • Employer does nothing 41%
  • Encouragement, tools and training 93%

S Brookings (Humlum, Vestergaard),

Top reads from the Library

All 32 top reads

The strongest sources the screening has found: original research and data that move the debate. Tier S is the top grade; how sources are graded.

  1. S

    Financing the AI Buildout (w35865)

    NBER (Columbia) · Paper · · 60 min original

    Cost per MW: about $11 million for the building, $2 million for power infrastructure, $28 million for GPU racks and IT; a 200 MW campus costs about $8.2 billion, two-thirds of it IT equipment.

    $41 billion estimated cost of a 1 GW AI campus ($41 million per MW)

  2. S

    Workforce in Motion: Skills and Pathways to Future Jobs in the United States

    McKinsey Global Institute · Report · · 60 min original

    About 36 million US jobs see reduced demand by 2035; about 41 million are created elsewhere.

    about 11 million US workers (roughly 7% of the workforce) who would need to leave their occupations by 2035 in the base case

  3. S

    The AI-as-Normal-Technology view of loss of control incidents

    Kapoor and Narayanan · Practitioner post · · 55 min original

    The Hugging Face incident occurred while OpenAI had monitoring for most internal use but not for evaluations, and the evaluation used a different harness and prompts than production Codex.

    >100x reduction in propensity to compromise out-of-scope infrastructure when OpenAI used its production Codex harness and prompt

  4. S

    The Economics of Transformative AI

    NBER · Paper · · 600 min original

    Value and growth: Agrawal, Gans and Goldfarb on "Genius on Demand"; Benjamin F. Jones on AI in research and development; Coyle and Poquiz on measuring AI in national statistics.

  5. S

    Economic Scenarios for Transformative AI

    Anthropic Institute · Paper · · 75 min original

    Modest: AI touches 4% of task instances by 2030; GDP +1.6%, growth 2.4% a year instead of 2%, labor share 60.0% to 59.4%, unemployment 3.9% against a normal 3.8%.

    +8.3% GDP above the no-AI path in 2030, substantial scenario; the median US adult's answers imply +8.6%

What the Library is made of

179 sources from 105 organisations, in 3languages.

By type

  • Report58
  • Paper53
  • Article40
  • Survey10
  • Talk or podcast10
  • Other8

By language

  • English158
  • Russian13
  • Chinese8

Browse all sources

A screening method I built runs every week through papers, reports, talks and posts, keeps the strongest, summarises each in English and grades it S to C. Most pieces here start from one of them.

Open the Library

Exhibits from the reading

More figures on Numbers

Figures as the sources published them, framed with source and date. Each opens full size.

Exhibit 1 · Report · Epoch AI

Key findings, as published

Screenshot of the Epoch AI report page "Can AI automate Epoch?", 8 October 2026, with the key findings list. Highlighted: models struggle to pick up Epoch's standards even with ample reference material, missing implicit conventions such as visual style and audience topics.
Highlighted: models miss the implicit conventions of the house. Source: Epoch AI, Can AI automate Epoch?, 8 Oct 2026. Page capture. A Library entry Original

Exhibit 2 · Paper · NBER w33777

The new AI tasks, by occupation

Stacked bar chart of new AI-related task types by occupation, from IT support to financial advisors: integration, content drafting, ethics and compliance, quality review, data insights, ideation. HR professionals rank third on integration work.
Integration work leads in IT support and software. HR professionals come third. Source: Humlum and Vestergaard, NBER w33777, revised March 2026, Fig. 4. S Same authors in the Library Original

Exhibit 3 · Paper · NBER w33641

With AI, commercial and R&D staff propose similar solutions

Two density plots of solution technicality for commercial and R&D professionals working alone. Without AI the distributions differ, R&D more technical; with AI they almost overlap.
Without AI the two functions answer differently. With AI the gap nearly closes. Source: Dell'Acqua et al., The Cybernetic Teammate, NBER w33641, April 2025, Fig. 6. Not filed in the Library Original

Long read

Long reads (4)

My long-form writing, shown with the drawings and builds behind it.

Ivan's one-page diagram 'Graph engineering': a frame feeds project facts, external research, stakeholder context and a counter-case into an evidence check, three candidate stories, a storyboard, parallel review and release, with human gates marked where claims or layout fail.
My own diagram of the work graph behind one piece of writing, published 8 Sep 2026.

Essay·7 min read

The Specification Problem: Why We Are Paying a 40% "Rework Tax" on AI

Developers built Spec-Driven Development to control AI coding agents. The same principles can restructure analysis, planning, reporting and decision work.

The test: where do the corrections to your last AI-made deliverable live now? If the answer is "inside the chat", the next run starts without them.

Conventional AI-assisted workSpecification-Driven Work
Add files and corrections to one growing conversationResearch first, then carry forward only relevant context
Judge quality after the draft appearsDefine acceptance before execution
Lose the accepted path when the chat closesRetain the method for the next run or colleague
Synthetic decision brief, first version, for a fictional company, built by Ivan.
Before. My build, synthetic data, shown July 2026.
The same synthetic decision brief after correction, titled Capacity without net headcount growth, with a recommendation, an evidence chart and investigation questions.
After correction. My build, synthetic data, July 2026.

A decision brief I built on synthetic data, before and after its rules were corrected.

Read the essay

This week in AI, summarised

All weekly summaries

Each week, what public sources reported on AI and work and what it means, with every source linked.

3-10 October 2026·8 stories

The week in AI, 3-10 October 2026

  1. Checking is now the scarce resource

    OpenAI published 722 machine-generated maths manuscripts and withdrew three within days; a maths site froze proof claims after 291 arrived in one day; open-source code output is up 183% while review comments per line fell by almost two thirds.

  2. Agents that act on the open web are producing incident reports

    Anthropic, the Wikimedia Foundation and an Australian parliamentary hearing each documented agents doing things nobody asked for: injection attacks, form submissions, proxy tricks.

  3. The harness buys cost, mostly

    Several independent measurements this week found that the scaffolding around a model changes the bill far more than the result, and two large companies reported that small open models with a validator beat frontier models on cost by one to two orders of magnitude.

Read the summary

Exhibit

The tasks worked. The job did not.

Epoch AI report page titled Can AI automate Epoch?, dated 8 October 2026, listing three key findings. The second, on models missing implicit conventions such as visual style and audience interests, is highlighted.
Can AI automate Epoch? Key findings section; highlight added by me. Source: epoch.ai