Michael Kilty · Fractional Head of AI · Across Europe

Get AI doing real work in your business.

I find the jobs AI can take off your team's desks, like proposals, document checks and monthly reports. Then I put the tools in place and stay until people use them every day. You don't need a tech department to start.

Book a 15-minute call → How a first project runs

Free. No slides, no preparation. Just bring one task.

AI training compute since 2010Higher dots took more computing to train · log scale, marked against GPT-3

The interactive chart needs WebGL, which this browser has switched off. The record-setting models are listed below.

The largest training run in this data is estimated at about 3,000 times GPT-3's training compute. That doesn't tell you whether a tool will work for your task. I help you find out, with your team, on real work.

Every dot is a real AI model with a published training-compute estimate. Tap a dot for its story, or use the arrow keys. Data: Epoch AI (CC BY 4.0, filtered and merged) and models.dev (MIT licence, © 2025 models.dev), checked daily; last check 25 September 2026.

See the 32 record-setters as a list

Each model below used more training compute than any earlier model in this data. Compute is shown relative to GPT-3 (2020).

  1. 2010-05Feedforward NN · University of Montreal / Université de Montréal · ≈1/910,000,000 of GPT-3 · Digit recognition
  2. 2010-06iCCCP · Massachusetts Institute of Technology (MIT) · ≈1/300,000,000 of GPT-3 · Object detection
  3. 2010-09Pooling CNN (NORB) · University of Bonn · ≈1/220,000,000 of GPT-3 · Image classification
  4. 2010-09RNN LM · Johns Hopkins University · ≈1/5,900,000 of GPT-3 · Language generation
  5. 2012-06Dropout (ImageNet) · University of Toronto · ≈1/1,100,000 of GPT-3 · Image classification
  6. 2012-07Unsupervised High-level Feature Learner · Google · ≈1/520,000 of GPT-3 · Image classification
  7. 2013-01DistBelief NNLM · Google · ≈1/120,000 of GPT-3 · Semantic embedding
  8. 2014-06SPPNet · Microsoft · ≈1/93,000 of GPT-3 · Image classification
  9. 2014-09VGG19 · University of Oxford · ≈1/29,000 of GPT-3 · Image classification
  10. 2014-09VGG16 · University of Oxford · ≈1/26,000 of GPT-3 · Image classification
  11. 2014-09Seq2Seq LSTM · Google · ≈1/5,600 of GPT-3 · Translation
  12. 2014-12SNM-skip · Google · ≈1/1,100 of GPT-3 · Language generation
  13. 2015-10AlphaGo Fan · DeepMind · ≈1/830 of GPT-3 · Go
  14. 2016-01AlphaGo Lee · DeepMind · ≈1/170 of GPT-3 · Go
  15. 2016-09GNMT · Google · ≈1/48 of GPT-3 · Translation
  16. 2018-05ResNeXt-101 32x48d · Facebook · ≈1/36 of GPT-3 · Image classification
  17. 2019-09Megatron-LM (8.3B) · Nvidia · ≈1/35 of GPT-3 · Language generation
  18. 2019-09Megatron-BERT · Nvidia · ≈1/14 of GPT-3 · Language generation
  19. 2019-10T5-11B · Google · ≈1/10 of GPT-3 · Text autocompletion, Language generation
  20. 2019-10AlphaStar · DeepMind · ≈1/3 of GPT-3 · StarCraft
  21. 2020-01Meena · Google Brain · ≈1/3 of GPT-3 · Text autocompletion, Chat
  22. 2020-05GPT-3 175B (davinci) · OpenAI · ≈ GPT-3 · Text autocompletion, Language generation
  23. 2021-08Jurassic-1-Jumbo · AI21 Labs · ≈ GPT-3 · Language generation, Chat
  24. 2021-09FLAN 137B · Google Research · ≈6.5× GPT-3 · Language generation, Question answering
  25. 2022-03GPT-3.5 (davinci-002) · OpenAI · ≈8.1× GPT-3 · Language generation
  26. 2022-06Minerva (540B) · Google · ≈8.7× GPT-3 · Quantitative reasoning, Mathematical reasoning
  27. 2023-03GPT-4 (Mar 2023) · OpenAI · ≈66× GPT-3 · Language generation, Question answering
  28. 2023-12Gemini 1.0 Ultra · Google DeepMind · ≈160× GPT-3 · Language generation, Visual question answering
  29. 2025-02Grok 3 · xAI · ≈1,100× GPT-3 · Chat, Language generation
  30. 2025-02GPT-4.5 · OpenAI · ≈1,200× GPT-3 · Language generation, Question answering
  31. 2025-07Grok 4 · xAI · ≈1,600× GPT-3 · Language generation, Question answering
  32. 2026-09GPT-6 Astra · OpenAI · ≈3,200× GPT-3 · Language generation, Question answering

Latest AI releases

Checked daily · last check 25 September 2026

  1. Qwen 3.8 Max PrimeAlibaba Not on the chart: training compute unavailable in our sources. Source: models.dev.
  2. Claude Opus 5.5Anthropic Not on the chart: training compute unavailable in our sources. Source: models.dev.
  3. GPT-6 LunaOpenAI Not on the chart: training compute unavailable in our sources. Source: models.dev.
  4. GPT-6 SolOpenAI Not on the chart: training compute unavailable in our sources. Source: models.dev.
  5. MiMo-V2.6-FlashXiaomi Not on the chart: training compute unavailable in our sources. Source: models.dev.
All 83 releases in the last 90 days
  1. MiMo-V2.6-ProXiaomi Not on the chart: training compute unavailable in our sources. Source: models.dev.
  2. Grok 4.7xAI Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  3. MiMo-V2.6-Pro-UltraSpeedXiaomi Not on the chart: training compute unavailable in our sources. Source: models.dev.
  4. Viv FastVivgrid Not on the chart: training compute unavailable in our sources. Source: models.dev.
  5. Step 5 PreviewStepFun Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  6. ParetoUnbiased Not on the chart: training compute unavailable in our sources. Source: models.dev.
  7. Qwen3.8 Omni FlashAlibaba Not on the chart: training compute unavailable in our sources. Source: models.dev.
  8. Arrow 2QuiverAI Not on the chart: training compute unavailable in our sources. Source: models.dev.
  9. Arrow 2 TelosQuiverAI Not on the chart: training compute unavailable in our sources. Source: models.dev.
  10. Kimi K2.8 PreviewMoonshot AI Not on the chart: training compute unavailable in our sources. Source: models.dev.
  11. DeepSeek V4.1 FlashDeepSeek On the chart. Source: Epoch AI.
  12. ChatGPT Images 2.5OpenAI Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  13. GPT-Image-2.5 FlareOpenAI Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  14. GPT-Image-2.5 SunburstOpenAI Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  15. MiniCPM5-2BOpenBMB Not on the chart: training compute unavailable in our sources. Source: models.dev.
  16. GPT-6 AstraOpenAI On the chart. Source: Epoch AI.
  17. OpenEvidence DarwinOpenEvidence Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  18. WeatherNext 3Google DeepMind On the chart. Source: Epoch AI.
  19. Gemini 3.8 FlashGoogle DeepMind Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  20. Muse Spark 1.3Meta AI Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  21. Qwen3.8 Max 0902Alibaba Not on the chart: training compute unavailable in our sources. Source: models.dev.
  22. Claude Fable 5.1Anthropic Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  23. Hy4 previewTencent Not on the chart: training compute unavailable in our sources. Source: models.dev.
  24. Tencent Hy4 previewTencent Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  25. Ling 3.0 Flash FininclusionAI Not on the chart: training compute unavailable in our sources. Source: models.dev.
  26. Qwen3.8 Flash NextAlibaba Not on the chart: training compute unavailable in our sources. Source: models.dev.
  27. Gemini 3.5 Transcribe LiveGoogle Not on the chart: training compute unavailable in our sources. Source: models.dev.
  28. Qwen3.8 FlashAlibaba Not on the chart: training compute unavailable in our sources. Source: models.dev.
  29. DeepSeek V4 Flash Vision ExpDeepSeek Not on the chart: training compute unavailable in our sources. Source: models.dev.
  30. GLM-5.3-FlashZ.ai (Zhipu AI) On the chart. Source: Epoch AI.
  31. GEN-1.5Generalist Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  32. Ornith 1.5 35B A3BDeepReinforce Not on the chart: training compute unavailable in our sources. Source: models.dev.
  33. GLM-5.3Z.ai (Zhipu AI) Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  34. Qwen3.8 27BAlibaba Not on the chart: training compute unavailable in our sources. Source: models.dev.
  35. Gemini 3.7 FlashGoogle Not on the chart: training compute unavailable in our sources. Source: models.dev.
  36. Gemini Flash LatestGoogle Not on the chart: training compute unavailable in our sources. Source: models.dev.
  37. Toast 1Mixedbread Not on the chart: training compute unavailable in our sources. Source: models.dev.
  38. DeepSeek V4 Pro 0813DeepSeek Not on the chart: training compute unavailable in our sources. Source: models.dev.
  39. Grok 4.6xAI Not on the chart: training compute unavailable in our sources. Source: models.dev.
  40. GPT-5.6 CyberOpenAI Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  41. MAI-Code-1.1-FlashMicrosoft Not on the chart: training compute unavailable in our sources. Source: models.dev.
  42. Nemotron 3.5 Lightning 30B A3BNVIDIA Not on the chart: training compute unavailable in our sources. Source: models.dev.
  43. Muse Glimmer 30BMeta Not on the chart: training compute unavailable in our sources. Source: models.dev.
  44. Qwen3.8-2.4T-A95BAlibaba Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  45. Grok Imagine Image 2.0xAI Not on the chart: training compute unavailable in our sources. Source: models.dev.
  46. Motif-3Motif Technologies On the chart. Source: Epoch AI.
  47. Solar Pro 4Upstage Not on the chart: training compute unavailable in our sources. Source: models.dev.
  48. Muse Spark 1.2Meta AI Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  49. Grok Voice STT 1.0xAI Not on the chart: training compute unavailable in our sources. Source: models.dev.
  50. Sakana NamazuSakana AI Not on the chart: training compute unavailable in our sources. Source: models.dev.
  51. DeepSeek V4 Flash 0731DeepSeek Not on the chart: training compute unavailable in our sources. Source: models.dev.
  52. Grok Voice TTS 1.0xAI Not on the chart: training compute unavailable in our sources. Source: models.dev.
  53. K-EXAONE 2.0LG AI Research On the chart. Source: Epoch AI.
  54. Inkling SmallThinking Machines Not on the chart: training compute unavailable in our sources. Source: models.dev.
  55. A.X K2SK Telecom On the chart. Source: Epoch AI.
  56. Claude Opus 5Anthropic Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  57. Gemini 3.5 Flash LiteGoogle Not on the chart: training compute unavailable in our sources. Source: models.dev.
  58. Gemini 3.6 FlashGoogle Not on the chart: training compute unavailable in our sources. Source: models.dev.
  59. Gemini Flash-Lite LatestGoogle Not on the chart: training compute unavailable in our sources. Source: models.dev.
  60. Laguna S 2.1Poolside Not on the chart: training compute unavailable in our sources. Source: models.dev.
  61. Qwen 3.8 MaxAlibaba Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  62. Qwen3.8 Max PreviewAlibaba Not on the chart: training compute unavailable in our sources. Source: models.dev.
  63. Kimi K3Moonshot On the chart. Source: Epoch AI.
  64. InklingThinking Machines On the chart. Source: Epoch AI.
  65. Qwen3.7 FlashAlibaba Not on the chart: training compute unavailable in our sources. Source: models.dev.
  66. GPT-5.6 LunaOpenAI Not on the chart: training compute unavailable in our sources. Source: models.dev.
  67. GPT-5.6 SolOpenAI Not on the chart: training compute unavailable in our sources. Source: models.dev.
  68. GPT-5.6 TerraOpenAI Not on the chart: training compute unavailable in our sources. Source: models.dev.
  69. Grok 4.5xAI Not on the chart: training compute unavailable in our sources. Source: models.dev.
  70. GPT-Realtime-2.1OpenAI Not on the chart: training compute unavailable in our sources. Source: models.dev.
  71. Hy3Tencent Not on the chart: training compute unavailable in our sources. Source: models.dev.
  72. Tencent Hy3Tencent Not on the chart: training compute unavailable in our sources. Source: Epoch AI.
  73. Laguna XS 2.1Poolside Not on the chart: training compute unavailable in our sources. Source: models.dev.
  74. Claude Sonnet 5Anthropic Not on the chart: training compute unavailable in our sources. Source: models.dev.
  75. Gemini Omni Flash PreviewGoogle Not on the chart: training compute unavailable in our sources. Source: models.dev.
  76. Nano Banana 2 LiteGoogle Not on the chart: training compute unavailable in our sources. Source: models.dev.
  77. LongCat-2.0Meituan Inc On the chart. Source: Epoch AI.
  78. Solar Open2 250BUpstage On the chart. Source: Epoch AI.

On the chartNot on the chart: training compute unavailable in our sources

Experience27 years on the business side of technology
Track recordWarner Music · NBC Director of IT at Warner Music, Los Angeles. Olympic operations for NBC, Athens.
How I workOne person accountable, from the first call to your team using it every day

Does this sound like your company?

Most leaders I speak to aren't short of AI ideas. What they lack is someone to pick the right one and get it working.

The side project

Everyone is trying ChatGPT, and nothing has actually changed.

People use it on the side. The work, the costs and the results look the same as last year.

The queue

IT says six months. You need it this quarter.

Your department has a clear problem, and it keeps slipping down someone else's list.

The report on the shelf

You paid for a strategy. Nobody built anything.

The report was sensible. What you needed was something your team could use on Monday.

The admin trap

Your best people spend too much time on admin.

Quotes, reports, emails and document checks. Necessary work, but not what you hired them for.

The 45-second version

What it looks like when AI actually gets built.

No slides and no report on the shelf. One task at a time, running in your business, with a person checking every output. Here is the whole approach in 45 seconds.

Every point is on screen, so it works without sound too.

Book a free 15-minute call →

A first project takes about four weeks.

We start with one task, prove the value, and only then decide what comes next. You deal with me throughout, with no hand-off to a junior team.

Week 1

Find it

Half a day with the people who do the work. We pick a task that's worth fixing and realistic to fix, check we can reach the information it needs, and record how it's done today.

Weeks 2–3

Build it

I set the AI up on your own documents, templates and rules, and test it with your team while it takes shape.

Week 4

Hand it over

Your people use it on real work. We compare it with how it was done before, and you decide whether to do the next one.

What you have after week 4
  • A working tool, running on your own documents
  • A named owner in your team
  • Written instructions, and a person checking the output
  • Time per task before and after, including checking and corrections

Exact scope and timing are agreed in writing once we've checked what the task depends on.

Good places for a first AI project.

Typical first projects, by department. Each one keeps a person in charge of the result. Yours will depend on where the hours actually go, which is what the first week finds out.

Sales

Call notesa first-draft proposal on your templateyour salesperson checks and sends

Operations

Outgoing contracts checked against your current termsdifferences flaggedyour team decides

Customer service

An enquiry in Germana reply drafted from your own answerschecked before it's sent

Finance and admin

Invoices matched to purchase ordersmismatches flaggedpaid only after a person checks

People and HR

A staff question about a policya draft answer with the source pageyour HR team checks it

Marketing

One approved listingpages in three languageschecked before publication

Done for a client See how →

Built, not just advised.

Two very different businesses, one approach: learn how the work really moves, then direct AI to build what the people doing it need.

Yacht brokerage · Client · Live

A yacht brokerage: one approved listing, every output checked.

80+yachts listed
3languages
240+listing pages

Live sitemap count, 24 September 2026

A live website, a broker portal and two kinds of brochure, all fed from what the broker approved.

Read the case study →
Healthcare · My own venture · In development

A healthcare platform: one patient journey, four points of view.

4workspaces
3languages
500+automated tests

Build and test inventory, September 2026

Patients, doctors, clinics and pharmacies in one application. Every request has an owner at every step.

Read the case study →

“I don't know very much about technology, but I can certainly see the excellent result, and the care you've put into helping [the business] grow.”

Founder and brokerBoutique East Mediterranean yacht brokerage · translated from Greek
Why this matters now

Many old moats are now head starts.

AI has made much of what used to be slow and expensive to copy, especially software and repeatable workflows, far quicker to rebuild. That cuts both ways: a competitor can copy some of what you do faster, and you can build things that used to be out of reach.

Then

Walls that took years to build

Software that needed a development team. Systems only large companies could afford. A head start that took competitors years and money to close.

Now

One person, directing AI

I'm the example. With no development team, I built a live publishing system with 240+ pages in three languages for a yacht brokerage, and a healthcare platform, still in development, checked by 500+ automated tests.

What lasts

How fast you change

Relationships, reputation and hard-won know-how still count. So does how quickly your company turns a new idea into everyday work. That speed is what I help you build, one task at a time.

The 90-day test

A simple question to put to your business: if a well-funded competitor using AI set out to copy what you do, what could they plausibly rebuild in 90 days? In the opportunity review we work through it together, part by part, allowing for their data, integrations and quality bar. You get a list of what's easy to copy, what's hard to copy and why, and what you could now build yourself.

See the review →
Why a business person

I speak business. AI does the typing.

I'm a business leader, not a software engineer. I choose the work to improve, direct AI to build the tools, check the result and help your team use it.

I start from your results.

Every project begins with what it will save or earn, and which person in your team will own it.

I direct AI the way a good manager directs a strong team.

Clear briefs, checked work, delivered on time. I test each tool with your team against agreed examples. At handover, we name who checks its output and who maintains it.

I'll tell you when AI is the wrong answer.

Sometimes the fix is a better process or a simpler tool. You'll hear that from me before we agree to build anything.

Start small. Keep what works.

Three ways to work together. Each one is quoted in writing before anything starts, so you know the full cost up front.

A few days

AI opportunity review

Best if you're not sure where to start.

Which tasks are worth doing first, what each one needs, and roughly what it would cost. The plan is yours to keep, whoever does the work.

  • Ranked list of opportunities
  • Rough cost and time for each
  • The first project, scoped
One fixed fee, agreed before we start.
About four weeks · Usual starting point

First project

Best if you already have one task in mind.

One task handed over to AI, from choosing it to your team using it every day, with time per task compared before and after.

  • A working tool on your documents
  • A named owner and written instructions
  • Time per task, before and after
Fixed scope and fixed fee. No hourly billing.
Ongoing

Fractional Head of AI

Best if you need someone to lead AI over time.

Someone who owns AI for your company or department: priorities, suppliers, delivery and training, for a set number of days a month.

  • A monthly priorities review
  • Tool and supplier decisions
  • Team training and support
A monthly retainer sized to the days you need.
Book a free 15-minute call →

If there's a fit, you get a short written proposal with the scope, timing and fee. How we'd work together →

Michael Kilty
Tallinn · Estonia
Meet Michael

27 years on the business side of technology.

I'm Michael Kilty, founder of Arvanu. I've been a Director of IT at Warner Music Group in Los Angeles, worked on NBC's Olympic operations in Athens, founded and run a currency brokerage, and led marketing across a group of technology companies. Since 2021 I've led AI at Healthview AI, building software for regulated pharmacy work.

That mix is the point. I know what a board wants to see, what a department head is measured on, and what makes a team actually adopt a new tool.

1999–2001Director of IT, Warner Music Group
2003–2004IT senior consultant, NBC Olympics, Athens
2004–2009Founder, FXGreece brokerage
2017–2018Group CMO, ICE Technology Services
2021–nowHead of AI, Healthview AI
2024–nowFounder, Arvanu
More about me →LinkedIn ↗
Let's talk

Tell me one task your team hates doing.

In a free 15-minute call, we'll work out whether AI can take it on and what a sensible first step would be. If the answer is no, I'll say so.

Book a 15-minute call → michael@arvanu.com

Free · 15 minutes · you'll talk to me, not a sales team.

Arvanu, in 45 seconds

Ready when you are.Book a free 15-minute call