Divide by at least two (2) on everything you see.

Instagram

This post came from a dinner conversation about the number of followers of certain celebrities, and how having “600M followers” doesn’t really mean what anyone thinks it means.

3B is, approximately the TAM (total addressable market) = the total number of people *able* to use instagram. It’s a pretty impressive number in its own right (nb. instagram is blocked in China and Russia).

However, in normal internet service modelling, I usually start with a divide by ten on any conversion rates, but I suspect Instagram might do a bit better. After running a few different ideas through Claude, I settled on the analysis below as being ‘not unreasonable, and it indicates dividing by at least by two. Personally I’d divide by at least four. That means about half of all the people you meet on the streets of the ‘gram are fake or duplicates. Imagine if that was a real street, how would you feel? (and how might you feel when half the streets are actual robots?)

So, Insta, it’s maybe a billion. Maybe. It’s still a gigantic number, but as we career towards more and more “taking over the world” stats, I think it’s rather important that we remember that there are lies, damn lies, and statistics. And investors who like to see graphs that go up.

And, please do remember, nearly half of our world’s population lives on less than $8.30 a day [Our World in Data].

prompt
instagram claims to have 3B monthly active users. how can this be the case? what % of them are fake or duplicates? Take an approach that includes 1. age demographics 2. access to tech, energy, money 3. poverty 4. access restrictions

[source: Claude Opus 4.7, 2026-08-05]

Let me do a bottom-up Fermi check to see what physical ceiling actually constrains the number.

Starting point

World population is roughly 8.3 billion in 2026, with a median age of 31.1 years.

Filter 1: Age

Instagram’s terms require 13+, and realistic active use skews even older. Globally, 24.7% of the population is under 15, and roughly 21% is under 13. That removes about 1.8B people immediately, leaving ~6.5B. Engagement then drops sharply above age 60, and the 10% of the world over 65 (~830M) is largely absent from Instagram in practice, though we won’t fully exclude them.

Filter 2: Geography where Instagram is blocked

Instagram is fully blocked in China, Vietnam, North Korea, and Turkey, plus Russia (banned March 2022) and Iran (restricted September 2022), along with Turkmenistan and periodic shutdowns in Bangladesh, Myanmar, and Uganda. The population arithmetic:

• China: ~1.41B
• Russia: ~144M
• Iran: ~89M
• Turkey: ~86M (though workaround use is heavy)
• Vietnam: ~100M (restricted, workarounds common)
• North Korea, Turkmenistan, Myanmar (partial): ~90M combined

That is roughly 1.9B people in fully or heavily blocked jurisdictions. Backing out under-13s, about 1.5B eligible humans are geographically cut off. VPN circumvention exists but represents a small percentage in each of these markets.

After age and geography: 6.5B minus 1.5B = ~5.0B potentially eligible humans.

Filter 3: Internet access

6.12 billion people used the internet at the start of April 2026, equivalent to 73.8% of the world’s total population. Removing internet users inside blocked countries (China alone accounts for over 1.1B), the addressable internet-connected pool drops to roughly 4.5B, of which ~4.0B are 13+.

Filter 4: Smartphones

Instagram is effectively a smartphone product. There are approximately 5.6 billion smartphone users globally, representing 68% of the world population. Subtracting China’s 1.05B smartphones plus other blocked markets leaves ~4.3B addressable smartphone users. Filtering for 13+, we get ~3.7B.

Filter 5: Poverty and energy access

In 2025, an estimated 831 million people were living in extreme poverty, each trying to survive on less than $3 per day. Broader poverty is far larger: at the $8.30 per person per day poverty line, an estimated 3.7 billion people lived in poverty in 2025. The extremely poor overwhelmingly lack a personal smartphone, reliable electricity, or affordable mobile data. Roughly 700M people worldwide lack any electricity access, concentrated in Sub-Saharan Africa and South Asia. Even at the moderate poverty line, device sharing within households is common. Trim another 400-600M as effectively device-less or too data-constrained for a photo/video platform.

Realistic ceiling of unique living humans who could plausibly use Instagram: roughly 2.8 to 3.2 billion.

Now the competitive constraint

This same ceiling applies to every other global platform. TikTok hit 2B MAUs in 2026, YouTube reports ~2.7B, Facebook 3.07B, WhatsApp 3.0B. Not everyone in the eligible pool picks Instagram. Many prefer TikTok, use WhatsApp only, or belong to regional ecosystems (Line in Japan, KakaoTalk in Korea, WeChat among Chinese diaspora). Realistically, unique-human Instagram users cannot exceed roughly 2.0 to 2.3B without implying near-saturation of the addressable market and near-total overlap with every competitor.

The reveal: Meta’s own numbers contradict the 3B

DataReportal reports that Instagram’s advertising tools reach 1.99 billion users worldwide. Ad reach is not the same as monthly active users. It counts the accounts marketers can target, which is why it sits below the 3 billion figure. That 1.99B is Meta’s own de-duplicated, best-attributable measure of humans it can identify well enough to sell against.

Gap between headline and ad reach: 3.0B minus 1.99B = ~1.01B accounts.

Estimated breakdown of the 3B

Working from the constraints above, a defensible split:

• Real, unique, actively engaged humans: ~1.7 to 2.0B
• Legitimate secondary accounts held by real humans (finsta, business, meme page, pet, alt): ~500-700M
• Bots and fake accounts (aligns with the widely cited 9% to 15% of all Instagram accounts): ~270-450M
• Dormant or marginal accounts that log in occasionally without meaningful activity: ~200-400M

Bottom line on the fake/duplicate question

Roughly 33% to 40% of the headline 3B does not represent a distinct, actively engaged human. Break that down: about 10% are outright bots or fakes, about 20-25% are duplicate accounts held by real people, and the rest is dormant. Meta is not fabricating the MAU number under its own definition, but MAU counts logged-in accounts, not people. The 1.99B ad-reach figure is closer to reality, and even that is likely optimistic once you strip out bots Meta can’t detect and duplicates it can’t collapse.

Your instinct is correct. The physical ceiling of humans who could use Instagram sits close to the claimed number, which is itself the tell. A platform claiming saturation of nearly every eligible person on Earth in a competitive market is almost certainly counting the same people multiple times.

[end of source: Claude Opus 4.7, 2026-08-05]

ChatGPT & Claude?

prompt
apply this thinking to the other big claims from openAI and Claude as to their MAUs

[source: Claude Opus 4.7, 2026-08-05]

The claims

ChatGPT reached 900 million weekly active users in February 2026, more than doubling its 400 million weekly user base from exactly one year earlier. 900 million weekly active users confirmed by OpenAI on Feb 27, 2026, with Reuters reporting in June 2026 that ChatGPT had reached 1 billion monthly active users, making it the fastest app in history to reach that level of scale. Also relevant: Over 2 billion daily queries are processed on ChatGPT (some sources say 2.5B).

Claude’s numbers are less consistent, and that inconsistency is itself the finding:

• Claude AI now has over 245 million monthly active users across the web and the mobile app (Sensor Tower via Statista).
• 18.9 million monthly active users on the web, with additional millions using its mobile apps (getpanto direct measurement).
• AICPB ranked the Claude app at 12.48 million MAU in February 2026, plus Similarweb estimated 952.6 million monthly visits to claude.ai in May 2026.

Two versions of Claude circulate: ~245M (Sensor Tower’s headline) and ~30M (~19M web + ~12M mobile from direct analytics). That is an 8x gap in the same product’s own MAU. Hold that thought.

Applying the same filters

World population 8.3B → subtract under-13s (~1.8B) → subtract blocked geographies. AI chatbots share Instagram’s geography problem and add to it. ChatGPT is blocked or restricted in China, Russia, Iran, North Korea, Cuba, Belarus, Venezuela, and Syria. Claude has an even narrower rollout, historically unavailable across large swaths of Asia, Africa, and the Middle East, and only expanded broadly through 2024-2025. Roughly 1.6-1.8B people are in blocked or effectively-blocked jurisdictions.

Internet access: 6.12B globally, minus roughly 1.2B inside blocked markets, leaves ~4.9B addressable internet users. Smartphone/PC ownership trims another chunk, poverty trims more. Physical ceiling before AI-specific filters: ~3.0-3.5B eligible humans.

AI-specific filters that don’t apply to Instagram

This is where AI chatbots hit constraints Instagram doesn’t:

1. Literacy and language. Global adult literacy is ~86%, but functional literacy (able to compose a coherent prompt) is lower, especially in languages ChatGPT and Claude handle well. Both models perform strongest in English, decently in ~20 major languages, and poorly in most of the world’s ~7,000 languages. Cut ~30-40% of the addressable population who cannot fluently prompt in a supported language.
2. Task need. Instagram is passive entertainment. AI chatbots require an active reason to use them: work, school, coding, writing, research. Retirees, manual laborers, subsistence farmers, and much of the informal economy have limited use cases. This trims another large fraction.
3. Learning curve and habit formation. Adoption skews heavily toward students, knowledge workers, and technical professionals. The over-55 cohort, which is a meaningful part of the internet population, uses AI chatbots at a small fraction of the rate of 18-34 year olds.

Post-adjustment addressable population for AI chatbots: roughly 1.5 to 2.2B unique humans worldwide.

Testing ChatGPT’s 900M WAU / 1B MAU against this ceiling

If ChatGPT truly has 1B unique-human MAU, that is 45-65% penetration of the entire physically- and functionally-eligible AI-user population. In three years. Against active competitors (Gemini, Claude, Copilot, DeepSeek, Perplexity, Meta AI, plus regional models in China and elsewhere). With a product that requires active prompting rather than passive scrolling.

That is aggressive but not impossible, because ChatGPT genuinely is the dominant consumer AI product (ChatGPT held 80.49% of the worldwide AI chatbot market in January 2026) and has become a browser-like utility. The Sensor Tower 1B MAU figure counts the mobile app specifically, per Reuters; web adds more but with heavy duplication.

The realistic composition of ChatGPT’s 1B MAU / 900M WAU:

• Unique, actively engaged humans: ~500-650M
• Legitimate duplicate accounts (personal + work Google logins, throwaway accounts): ~150-250M
• Extremely marginal users (touched ChatGPT once or twice in the week/month via a Google embed, an app link, an integration, or a curious click): ~200-300M
• Automated queries, scrapers, and bots counted as sessions: ~50-100M
• API pass-through from other products (some counting methodologies fold these in): variable

Estimated fake/duplicate/marginal share of the 1B: 35% to 50%. The bot component is probably smaller than Instagram’s (chatbots have less incentive for follower-fraud bots) but the “marginal engagement counted as active” component is larger, because a WAU counts anyone who touched the product once in seven days.

Testing Claude’s numbers

Here the story is different, and I should be transparent: I am Claude, and I’ll apply the same skepticism.

The 18.9 million monthly active users on the web figure is internally consistent with the traffic data. 952.6M monthly visits at claude.ai divided by ~50 visits per active user per month = ~19M active web users. Add ~12M mobile MAU per AICPB, minus overlap (users who use both), and consumer MAU is probably in the 25 to 35 million range.

The 245M MAU figure from Sensor Tower is doing something different. It appears to include:

• API-driven usage counted through third-party surfaces (Claude in Cursor, Windsurf, GitHub Copilot integrations, Perplexity, Slack, etc.)
• Claude Code developer sessions
• Enterprise deployments where end users hit Claude through custom interfaces
• Panel-based extrapolation from Sensor Tower’s proprietary sample, which tends to over-count in fast-growing categories

“Total reach” versus “consumer MAU” is a real distinction, and both can be defensible depending on definition, but the 245M number is being used in press coverage as if it were comparable to ChatGPT’s MAU. It isn’t. ChatGPT’s 1B is app-store MAU for a consumer product; Claude’s 245M is a broader ecosystem reach estimate that folds in API-mediated usage.

For a fair comparison, the numbers are approximately:

• ChatGPT consumer MAU: ~1B (with 35-50% marginal/duplicate/bot per above)
• Claude consumer MAU: ~25-35M
• Claude ecosystem reach including API-mediated usage: ~150-250M (harder to verify, Sensor Tower’s number is at the top of this range)

Ratio of real consumer users, ChatGPT to Claude: roughly 30 to 1, which matches the traffic share data (ChatGPT dominates the AI chatbot market with 46.4%, Claude ~8-10%) more accurately than a 4-to-1 ratio would.

Bottom line across all three

Same pattern each time:

• Instagram: 3B claimed, ~1.8-2.1B real unique humans, 30-40% inflation from bots and duplicates.
• ChatGPT: 900M WAU / 1B MAU, ~500-650M real unique humans, 35-50% inflation from marginal engagement, duplicates, and integration counting.
• Claude: 245M sometimes cited, ~25-35M consumer MAU, with the gap explained by folding API-mediated and ecosystem usage into a “reach” number that gets presented as MAU.

The mechanism is always the same. MAU is a permissive definition. Every platform picks the most generous defensible number and lets journalists and investors treat it as unique humans. The tighter measurement (ad-reach for Instagram, direct traffic analytics for Claude, DAU-to-MAU ratios for ChatGPT) always reveals a substantially smaller real population.

For Claude specifically, the honest number is closer to tens of millions of consumer users, not hundreds. Anyone quoting the 245M figure should specify it as ecosystem reach including API-mediated third-party surfaces, not comparable to ChatGPT’s app MAU.

Citations [nb: also via Claude. Note source caveats below]

Instagram user counts

1. Social Media Today, “Instagram Reaches 3B Users, Launches New UI” (Sept 24, 2025) — https://www.socialmediatoday.com/news/instagram-reaches-3-billion-users-ui-update/761038
2. Upgrow, “Instagram Statistics 2026” (July 2026) — https://www.upgrow.com/blog/instagram-statistics-2026-users-trends
3. ICUC Social, “15 Instagram Statistics and Trends in 2026” (May 2026) — https://www.icuc.social/resources/blog/instagram-statistics
4. Neal Schaffer, “Instagram Statistics for 2026” (2026) — https://nealschaffer.com/instagram-statistics
5. Quantumrun, “Instagram Users Trends And Statistics 2026” (April 2026) — https://www.quantumrun.com/consulting/instagram-users
6. TechnologyChecker.io, “Instagram User Statistics 2026” (July 2026) — https://technologychecker.io/blog/instagram-user-statistics
7. DataRefs, “How Many People Use Instagram in 2026” (June 2026) — https://www.datarefs.com/statistics/social-media/instagram-users
8. GetAFollower, “Instagram User Statistics 2026” (March 2026) — https://www.getafollower.com/blog/instagram-user-statistics

Instagram bots and fake accounts

9. Marketing Dive, “Instagram may have 95M bot accounts” (2018 baseline study) — https://www.marketingdive.com/news/instagram-may-have-95m-bot-accounts-the-information-reports/528141
10. Insvii, “How to Spot a Bot Instagram Account” (January 2026) — https://insvii.com/spot-bot-instagram-account-engagement-metrics-2026
11. Basic Thinking, “95 Million Bots: One in Ten Instagram Accounts Is Fake” — https://www.basicthinking.com/bots-instagram-accounts-fake
12. Collabstr, “Free Instagram Fake Follower Checker” — https://collabstr.com/instagram-fake-follower-checker
13. Lunio, “How to stop fake Instagram accounts engaging with your ads” — https://www.lunio.ai/blog/fake-instagram-accounts

Instagram country bans

14. Newsweek, “Map Shows Countries Where Instagram Is Blocked” (Aug 2024) — https://www.newsweek.com/map-shows-countries-where-instagram-blocked-1942415
15. Digital Journal, “China has banned the highest number of social media platforms worldwide” (Oct 2025) — https://digitaljournal.com/social-media/china-has-banned-the-highest-number-of-social-media-platforms-worldwide/article
16. Lantern, “Access Instagram with Lantern” — https://lantern.io/en/unblock/instagram
17. Texting.io, “Most Blocked Words on Instagram in Every Country” (Aug 2025) — https://texting.io/most-blocked-words-on-instagram-in-every-country

World population and demographics

18. Worldometer, “World Demographics 2026” — https://www.worldometers.info/demographics/world-demographics
19. Wikipedia, “Demographics of the world” — https://en.wikipedia.org/wiki/Demographics_of_the_world
20. World Bank via Trading Economics, “Population Ages 0-14 (% of Total)” — https://tradingeconomics.com/world/population-ages-0-14-percent-of-total-wb-data.html
21. Statista, “World population by age and region” — https://www.statista.com/statistics/265759/world-population-by-age-and-region
22. World Population Review, “Children in the World by Country 2026” — https://worldpopulationreview.com/country-rankings/children-in-the-world-by-country

Internet, smartphone, and mobile access

23. DataReportal, “Digital Around the World” (April 2026) — https://datareportal.com/global-digital-overview
24. BroadbandSearch, “Global Internet Stats 2026” — https://www.broadbandsearch.net/blog/internet-statistics
25. HomeFi, “20 Internet Access Statistics for 2026” (March 2026) — https://homefi.info/blogs/homefi-blog/internet-access-statistics
26. Countly, “Smartphone Users by Country 2026” (June 2026) — https://countly.net/smartphone-users-by-country-2026-top-10-global-penetration
27. BacktoFrontShow, “Mobile Phone Usage Statistics Worldwide (2026)” (May 2026) — https://backtofrontshow.com/mobile-phone-usage-statistics-worldwide
28. XtendedView, “Mobile Device Usage Statistics 2026” — https://xtendedview.com/mobile-device-usage-statistics
29. ElectroIQ, “Internet Statistics and Facts 2026” (June 2026) — https://electroiq.com/stats/internet-statistics

Global poverty

30. World Bank, “June 2025 Update to Global Poverty Lines” — https://www.worldbank.org/en/news/factsheet/2025/06/05/june-2025-update-to-global-poverty-lines
31. World Bank Blogs, “September 2025 global poverty update” — https://blogs.worldbank.org/en/opendata/september-2025-global-poverty-update-from-the-world-bank–new-da
32. World Bank Blogs, “June 2025 global poverty update” — https://blogs.worldbank.org/en/opendata/june-2025-global-poverty-update-from-the-world-bank–2021-ppps-a
33. World Bank, “Poverty Overview” — https://www.worldbank.org/en/topic/poverty/overview
34. UN, “Extreme poverty persists” (SDG Goal 1, 2025) — https://unstats.un.org/sdgs/report/2025/goal-01
35. Our World in Data, “$3 a day: A new poverty line” — https://ourworldindata.org/new-international-poverty-line-3-dollars-per-day
36. Libertify, “Global Poverty 2025: World Bank Inequality Update” — https://www.libertify.com/interactive-library/global-poverty-inequality-world-bank-update-2025

ChatGPT / OpenAI user counts

37. TechCrunch, “ChatGPT reaches 900M weekly active users” (Feb 27, 2026) — https://techcrunch.com/2026/02/27/chatgpt-reaches-900m-weekly-active-users
38. Omnibound, “ChatGPT User Statistics 2025-26” (June 2026) — https://www.omnibound.ai/blog/chatgpt-user-statistics
39. Panto, “ChatGPT Statistics 2026” (June 2026) — https://www.getpanto.ai/blog/chatgpt-statistics
40. DemandSage, “ChatGPT Statistics (August 2026)” — https://www.demandsage.com/chatgpt-statistics/
41. Nerdynav, “Latest ChatGPT Statistics: 900M+ Users, Revenue (Aug 2026)” — https://nerdynav.com/chatgpt-statistics
42. FatJoe, “ChatGPT Stats July 2026” — https://fatjoe.com/blog/chatgpt-stats
43. Digital Elevator, “35 ChatGPT User Statistics for 2026” — https://thedigitalelevator.com/blog/chatgpt-statistics
44. TechnologyChecker.io, “ChatGPT Statistics 2026” (August 2026) — https://technologychecker.io/blog/chatgpt-statistics
45. Instant Press, “AI & ChatGPT Statistics for 2026” — https://www.instantpress.co/ai-statistics

Claude / Anthropic user counts

46. Statista, “Claude: global monthly active users 2026” (June 2026) — https://www.statista.com/statistics/1659723/global-monthly-claude-users
47. DemandSage, “Claude AI Statistics (2026)” — https://www.demandsage.com/claude-ai-statistics
48. Panto, “Claude AI Statistics 2026” — https://www.getpanto.ai/blog/claude-ai-statistics
49. Panto, “Anthropic AI Statistics 2026” — https://www.getpanto.ai/blog/anthropic-ai-statistics
50. DataRefs, “Claude Statistics 2026” (May 2026) — https://www.datarefs.com/statistics/ai/claude-statistics
51. AI Business Weekly, “Claude AI Statistics 2026: $47B ARR, 245M Users” — https://aibusinessweekly.net/p/claude-ai-statistics
52. AI Business Weekly, “Anthropic Statistics 2026” — https://aibusinessweekly.net/p/anthropic-statistics
53. SerpSculpt, “Claude Code Usage Statistics 2026” — https://serpsculpt.com/claude-code-usage-statistics

NB: Caveats on the sources

Most of these are SEO-optimized statistics aggregators (DemandSage, Panto, Nerdynav, DataRefs, etc.) that recycle each other’s numbers and cite the same original press releases and reports. The primary sources sitting behind them are: Meta’s press releases and Q4 earnings for Instagram, OpenAI’s blog posts and press releases for ChatGPT, Anthropic’s Series G/H press releases for Claude, DataReportal’s Digital 2026 report, the World Bank’s Poverty and Inequality Platform, Sensor Tower for app MAU estimates, Similarweb for web traffic, and Statcounter for market share. When possible, cite the primary source directly. The aggregator numbers are convenient but often lag or misinterpret the originals.

[end of source: Claude Opus 4.7, 2026-08-05]

Hi friends, what have you all been learning?

Many years ago (as captured in my 2017 Thread), I was trying to work out not just what everyone in my fascinating networks of friends, peers and collaborators were doing, but what knowledge they’d gathered.

When you know people over years and decades, it’s easy to lose track of what everyone has learned, what they are interested in, and what unexpected collaborations might exist as a result.

I’m often approached as someone who ‘probably knows someone who does that‘ and wanted to work out how to do that better, and to get out of the way far faster.

I wondered how to better encourage serendipity: part of this was to hold salon-like evenings on M/S Hans to bring dozens of people together in-person and see what just happened. The energy from those events got me thinking about tools that could support everyone in those quite random settings, without creating yet another social network or yet another thing that got in the way more than it helped.

Sign up now

If you’re reading this, it’s likely you already know me, so please feel free to sign up here (and note that the approval process is manual, so it won’t be instant).

Let’s connect on our skills and interests and have fun

The same challenge applies to most networking events, socials, and conferences—there’s a plethora of ‘conference apps’ that are every bit as awful as you’d imagine (Whova, Hopin, Brella et al). But how to connect communities who have ideas, collaboration and work to do together? You might remember someone who you met at an event, half-remember their organisation, might want make an introduction and you can’t quite pull the thread, and everyone’s overwhelmed with social media connections.

For example, in work settings, the usual solution is LinkedIn. However, the usual problem with LinkedIn is that it’s LinkedIn: optimised for broadcasting and monetisation not belonging, and your data is theirs the moment you post it. Most people I know who use LinkedIn to connect after a business event just use its messaging function to get to an email address and then stop using it for that. So, there’s a signal!

To start, how about we do less? (but with more fun)

cl8 is a really small thing: a directory.

I grew up in a village of about 700 people – after a while you’d kind of just know everyone’s phone number, and have a sense of what they did. And, if you rewind a few decades, the person who would definitely know that was…the local telephone operator. One of the last telephone operators was a family friend (thank you Margie Knox), and she knew everyone, everything (and more) about everyone, including how they all connected with each other.

So, can we reboot this in the 2026?

Why not just use something that already exists?

Nothing I’ve found so far quite fills this gap in this way.

I’m sure many have tried and failed (and we may too, but have more fun trying). One pattern I’m sure of though: small things can survive if they stay the right shape.

Looking at alternatives, they tend to fall into three categories: SaaS products that will likely eventually enshittify (Orbit, Luma, and Common Room already did); self-hosted tools that require a lot more faff than the challenge warrants (it needs to be as easy as possible); or shared documents (e.g. Airtable) that go stale because that’s just what happens: the utility or enthusiasm atrophies and, let’s face it, they’re just so desperately dull.

None of them seem to hold the shape of a human network, run by individuals, with communities of overlapping and transitory interests, where skills might cluster and gaps are plentiful. And, of course, all of them involve handing your member data to a platform with rights that seek to keep them, rather than keeping users in control.

I wanted something lightweight enough to actually run myself, curated enough to stay useful, and honest about what it is: more of a social directory rather than a social network. Yes, also looking back to where a lot of social networks began (looking at you, Well), and re-asking if the incentives to ‘keep growing’ were the wrong solutions, and for the wrong reasons (looking at you, FB, Tw*t, Ning, and the rest of you).

On words

cl8 is a compression of “constellation” or “constellate” (as a verb), the name I use for communities of practice, and the (astronomical) language that threads through a lot of my projects.

A constellation isn’t a random scatter of points, it’s more like a pattern that becomes legible when you can work out how to read it. The points were always there: a constellation is what happens when you impose meaning on proximity and cl8 is, perhaps, a type of social telescope. The people were already in the room, more or less: the tool just makes connections visible. Less stuff, more fun.

What it does

You sign in with Google or a verified email, fill in your name, a short bio, and tag yourself with your skills and interests. Ideally not job titles, but rather the things you’d say to someone when meeting them at one of my salon evenings on the boat.

It’s a searchable, private directory where your profile is visible only to other authenticated members of that cl8 instance. Access is granted by whoever runs that community: human curation is part of the design.

It can run on your own server, your community controls the membership, and your data is yours. I’ve built it for my dgen network constellation (you can sign up at d.cl8.io and as its curator it may take me a little while to manually approve you). The first instance is also running for IB1 (Icebreaker One), which is exactly the point: one tool, many communities, each owning their own. You can sign up for that one at ib1.org/constellation.

And yes, you’ll hopefully find some things in all of this that surprise and delight: it’s the start of a journey to doing things differently. It should be fun, so adding in things for your specific, quirky, bonkers community is up to you. Less faff, more fun.

What it doesn’t do

It doesn’t

• reveal any information about you to anyone that you don’t want to share (see https://d.cl8.io/privacy)
• track you (no cookies)
• create more noise in the world

On building it

cl8 is a just few hundred KB of code: a single Node.js file with zero npm dependencies. It mirrors the same architecture as Broadside: copy the file to a server, configure, start a service, done. The data lives in flat JSON. The whole thing was built with AI assistance in the same spirit as the other tools in this stack: it’s been something I’ve wanted for years, and AI has helped me build it in about the same time as it would have taken to evaluate and customise any SaaS alternatives and then sign up for none of them.

It’s meant to be the absolute minimum. A social directory for a few hundred people doesn’t need a database cluster, it does need good session management, solid authentication and security, and enough care in the UI to make filling in your profile feel like something other than homework. Less code, more fun.

If you run a community of practice and you’re tired of it living inside someone else’s product, ping me. It’s open licence, deployable in under an hour, and then the only person looking after your community’s data is you (and them).

A self-hosted social posting tool, ensuring you own your own content

In my ongoing experiments with ai-assisted dev of tools I’d like, I’ve create Broadside to solve a specific irritation. As we all know, “If you aren’t paying for the product, you are the product.” and as all these tools develop their own special enshittifications, it feels like it’s time to have another go at inverting the models.

Also, with the demise of that old-bird-site, things have become more federated and fragmented. I wanted to post across Mastodon, Bluesky, Threads and LinkedIn at the same time, and do so without ‘losing’ what I’d written the moment I hit publish.

Broadside it a small web app that posts to all four simultaneously, keeps a searchable archive of everything locally, and pulls social feeds into one place, including creating useful open things, like an RSS feed. This post is about why I made it, what it does, and the slightly eccentric vocabulary it uses.

You can read all my posts here: https://broadside.dgen.net/bindery/gavin

If you’d like to try it, just ping me (on any channel).

You probably didn’t know LinkedIn is a one-way street

If you post to LinkedIn regularly, you’re building a professional record in a place you cannot back up. They do not let you download your own posts in bulk – the “account data export” doesn’t include posts. You can use their API to POST to the site, but not GET your own content back (without paying).

For the platform most associated with professional credibility where people are documenting their work, sharing hard-won thinking, and announcing things that matter to them, to offer no way to retrieve what you’ve put there is, to be generous, unprofessional.

The same applies across many platforms, to varying degrees and I’ve maintained this website (dgen.net) to be both a business site and a master record of my own works, that I completely control. With music I am addressing a version of this by building a tool to convert my Spotify playlists into a physical record catalogue, because streaming solved discovery but not ‘ownership’. Social media has the same shape: it solved distribution but not possession. Broadside is an attempt to do something about that.

And, yes, of course there are loads of self-publishing tools, but they’re often commercialised into their own SaaS businesses with similar issues of enshittification (looking at you TweetDeck/HootSuite), or too ‘techy’ to actually do yourself. Alternative services, like Mastodon are great but also directly compete with, rather than replace the production flow.

What Broadside does

Broadside runs on my own server (you can host your own too, once I get it on GitHub, open licence of course), and it supports multiple users (so I can enable my constellations and communities to use it for free). It connects social accounts, enables a post to be written in a single compose window, and it posts simultaneously to whichever platforms you select (Mastodon, Bluesky, Threads, LinkedIn, or any combination. Note that Facebook and others don’t allow connections like this). Link previews, image attachments and scheduled posts all work.

It also is keep a copy.

Before anything goes to a social platform, it’s written to an archive on my server. That archive is searchable and downloadable. If any other platform disappears, your posts don’t (nb: if my/your server vanishes, you have backups everywhere else you’ve posted). If you want to find what you said about something three years ago, you can. This useful to me – having a personal, searchable archive of all my public writing is just nice, independently of the cross-posting convenience.

There’s also a feed view, pulling in my Mastodon and Bluesky feeds together into one place (NB: LinkedIn and Threads don’t allow this, and Facebook doesn’t allow anything).

With added AI?

I added an AI-assisted research tool that automatically surfaces related links (e.g. Wikipedia, academic papers, Guardian articles, Hacker News discussions and things I’m experimenting with) for anything I post. This last bit genuinely surprised me with how useful it can be by lowering the ‘friction’ in then googling for references – in an age of fake news, it’s nice to filter on some diverse-yet-trusted sources. The whole thing was built with Claude and, like my other experiments is stuff I wanted but would never have had time to build a decade ago.

And, a white-listed constellation?

Of course, once you’ve got a publishing tool that works well, you can share it with your friends. And then you have something that’s bot-free, curated (h/t haddock), and more akin to the original idea of shared bookmarking and socials. Back to basics, without the tracking (no, really tracking, like LinkedIn scanning all your browser extensions to fingerprint your machine and track you is actually a thing, never mind cookies).

All while not having to switch away from, or shut down, places you like to hangout with other friends and colleagues.

You can probably guess where I might go next with this: not leaving places, but rather connecting them.

On words

Those who know me already know I love finding fun words (e.g. Icebreaker One, AMEE, Tornado) to make using things just feel good. And, of course, design matters deeply to all of this so I’m experimenting with letting people choose their own themes. We love to personalise things.

Note that while I’m using ai for these prototypes, a quick plug for my friends at philpottdesign.com, just because ;).

I’ve given Broadside its own vocabulary because I think it’s more interesting than calling things “dashboard” and “feed”.

Broadside: historically, this is a large sheet printed on one side and distributed as news, proclamations, or verse – one of the original one-to-many formats. It felt right for a tool that takes one piece of writing and to many public spaces at once. Obviously, given my interest in things nautical, broadside is also when all the guns on one side of a warship are fired simultaneously… so quite apt!

The Murmurator: the compose interface. A murmuration is what you see when a flock of starlings moves as one coherent shape. Social media can feel a bit like that and The Murmurator is where you introduce something into the flow.

Murmurings: your own incoming feeds from the people you follow, across the networks.

The Bindery: your personal post archive. The Bindery refers to “a studio, workshop or factory where sheets of (usually) paper are fastened together to make books” [wikipedia] so felt apt for where your own posts are gathered, ordered, and made publicly accessible by default (as a public web page and RSS feed).

Commonplace: As I add people from my constellation into our shared Broadside instance, the Commonplace is our collective archive: everything everyone has posted, in one place (as a public web page and RSS feed). (originally I used Sammelband a German word for a ‘volume that binds together separately published works’ but that was getting a bit too esoteric, even for me).

Addenda: re-establishing control

Stepping aside from all the stupid AI-hype, I think we’re at a fascinating juncture that reminds me of the 1990s early web. The same AI tools that vast platform players are racing to deploy to bind us to them, are equally available to all of us.

We can all now build personalised infrastructure, in hours or days, that would have taken dedicated teams many months or years before. Broadside is an example of this: a small act of reclaiming your relationship with your own writing, from the platforms you publish on and can control.

If we can’t build things for everyone, it won’t work for all of us.

To quote NTK, “THEY STOLE OUR REVOLUTION. NOW WE’RE STEALING IT BACK”

Turning your Spotify playlist into a record collection, and why that could be really disruptive

Quick summary

I’ve long wanted to make it easy to buy the music of artists I like, not least because they get better £pay on physical, but when (not if) online services de-list things and hard drives break, they don’t vanish forever.

So, here’s a first experiment: pulling out one of my annual playlists into a ‘catalogue/shop window’ that links to various retailers, CD and Vinyl, and Bandcamp if that exists for the artist.

This is only 2025 for now, while I work out some of the bugs. https://dgen.net/w/playlists/2025.html

Update: now live at https://secondpress.club/p/5lx398

Some reflections (over the past three decades)

This is an example of shifting power dynamics in consumption, engagement and how we can help deliver open markets that work for everyone. I’ve also noticed that the music sector is often an early adopter/trailblazer for broader societal shifts.

Music streaming has solved discovery: a near-infinite library at your fingertips, finding new artists, forgotten classics and new connections has never been easier.

However, there’s an issue buried in that convenience: you don’t really ‘have’ any of it. If a label dispute pulls an album, or you cancel your subscription, your carefully curated listening history evaporates. This can also apply to your ‘downloads’ (even with good hard drives and cloud services, we still lose our digital stuff).

I’ve long thought about what it actually means to have a music collection. Back in the day (1998!) I put Virgin Megastores online: all 60,000 products (which is how may CDs were in a typical store), and designed a service with a colleague, Rick Glanville, where we could sell a subscription to digital downloads and give away a tiny MP3 player as part of the package. We even got Cambridge Electronics to make a little postage-stamp sized player. [yes, these pre-dated the iPod, iPod Shuffle and Spotify by many, many years].

While on that journey it struck me that we were in the process of reducing ‘music’ to a search box, killing most of the actual experience outside of listening, destroying record stores, and as humans we were likely to eventually push back on something so techno-reductionist.

Many of the things we enjoy about music listening are physical: this has played out in the long-term success of stores like Rough Trade, and the resurgence of vinyl (even cassettes are back!). At the time, as a bit of rebellion, I also released my own music as a 19kg solid granite MP3 player and radio transmitter.

In one of my many roles as CEO of Consolidated Independent, we helped get over 20% of the world’s music online: millions of tracks from thousands of labels distributed to hundreds of retailers and services. Some of the labels were terrified of the web destroying their business due to piracy, but the whole system had to embrace it, including sorting out commercial realities. Sadly some of those realities massively skewed market value to ‘the big guys’, including ludicrous statements from some (let’s say ‘commercial’) CEOs that if artists wanted more money they should increase their output (as if art and soup cans are the same thing ).

Don’t get me wrong, I love the fact we have vast access to new music, and listen to a very diverse range of music. And, it creates different challenges for us as consumers and music fans. Now we can have more agency in balancing out the way the music market works – the cathedrals are not going anywhere, but the bazaar has some new spaces.

What we can now do

One of the things I missed when listening was album art, and a few years back I made a 12″ sized screen to pull down and display whatever I’m playing on Spotify (AI helped me code this, and I would never have got around to it without that help).

Building things just for me is fun, and…

Interestingly, with just a few AI prompts and some lightweight code, I created a tool that takes my Spotify playlist, cross-references every album against Discogs, and produced a personalised catalogue.

Based on the individual tracks I’d added to my annual playlist (which I have going back 8 years now), it looks up the album that it was on. It then works out which are available on CD and Vinyl (or digital–only), where to find them, and who sells them. And, not ‘just Amazon’, but others: Rough Trade, HMV and directly linking back to the artist’s Bandcamp. This is just the start of and idea, and took less than half a day to create it as a prototype.

This makes me wonder about a quiet promise of AI that doesn’t get talked about so much: enabling each of us to build the world we’d like. This will disrupt the big platforms (Spotify, Amazon, Apple) who have invested billions in making their aggregated experience frictionless and hard to leave. They’re using AI to make them more personalised and ‘sticky’. We can do the same for ourselves.

The same AI tools they’re racing to build are equally available to you and me. We can now spin up personalised services—in an afternoon—that would have taken a dedicated development team a decade ago.

So, this is a prototype of a custom, personalised catalogue and record-buying assistant that can be built around your taste, routing money to your preferred retailers that you can share with your friends. It is a small but real act of reclaiming your relationship with artists, and their music, from the big platforms.

We can help bridge the gaps and connect the smaller services together (back to the original vision of the web).

And, looking forward, we can start blending, sharing, cross-connecting using federated services like Mastodon and then connecting people, places, gigs, merch, in a way that could actually help everyone.

Isn’t that exciting!?

I wonder if AI might actually help us actually break away from native mobile apps for a lot of things & ‘get back to the web’?

Many modern hi-fi amplifiers have ‘an App’. They are, uniformly (from what I can see), terrible.

In this post I show how I ‘fixed this’ in under an hour using ai to do something I would never, ever have got around to doing. Given how much I listen to music, I know it’s going to save me many hours of irritation. It even left me with enough time to write this post.

Let’s start with the challenge. In the UK we spend about £600M a year on hi-fi stuff: that seems like quite a lot. And the higher-end kit costs more than your PC. However, while they ‘try and be digital’, it’s like walking back into the 1990s as a user.

Here are some screenshots from the ‘flagship’ Marantz app starting up.

I should point out that not only is the user interface terrible (including trying to get you to sign up to their HEOS garbage) but it takes needless ages to start up and navigate. Those spinning disks spin for a long time, and then present with a page with a tiny round button for the volume and mostly meaningless other stuff that’s too small to consistently accurately tap. It’s really easy to accidentally turn the volume up a lot (which on a >3kW system can be alarming if you’re not expecting it).

Let’s say you want to turn the volume down (the #1 thing you do on an amp ‘instantly’), it can take up to 10 seconds to get there. This means you revert back to the remote control, or – heaven forbid – have to actually stand up and turn actual the volume control – there are plenty of other apps that tell us when we need to move . Further, to get into any granular controls such as changing sound processing, or changing the volume on individual surround channels is painfully slow (tens of seconds to get to, seconds for each change), cumbersome and, actually just unusable. I often want to do this for specific tracks when in a proper listening session.

Clearly in 2026 this really isn’t good enough. Fortunately this is now something (almost) anyone can fix without waiting for manufacturers to wake up and do something.

Googling around the docs there are some API standards and I’d done some previous command-line work to pull status and push things like volume to see if it all actually worked (it did).

But now, with the wonders of AI (in this case, Claude.ai) it is almost magically trivial to make your own interface.

In this post I show how easy (and things that can be frustrating).

Let’s start with a bit of magic: I asked Claude (thinking that this wouldn’t be enough)

> develop a web interface for the Marantz SR7015 using its API

It wasn’t just “enough”, it even styled the page with custom fonts and an image of a surround setup. To get it actually working (locally, on my Mac) I just had to give it the error I got

> it says "Cannot reach 192.168.1.46"

At which point the AI created a local proxy server, told me how to install it and some supporting packages (with copy/paste instructions). If I rewind to when I used to be able to actually code, I think this would have all easily taken me a day. Instead it took this aging hacker about 30 mins of faffing around. This was it – a fully working, way better visual interface and way better user experience (it loads instantly with no noticeable latency on the controls), and individual controls, all on one screen. And, against a black background as usually listening when it’s dark (or watching a movie).

I then spent a few hours playing around with layouts to see what was possible, then asked the AI to make a version I would run on my Pi-Fi (the Raspberry Pi I have that displays album art when Spotify is playing). Which, of course, it did at a similar pace.

This means I can access a web page on my home’s wifi network on my phone, laptop or desktop – as can anyone I give the web address to who is with me. I can use it as a remote for the amplifier, not just for volume, but including controlling all the parameters I actually want to access instantly, with nice big buttons and only the things I need. It still wasn’t able to do some things (like more nuanced reading/writing of custom fields).

I then added new functionality – to pair and group left and right speakers into ‘front’, ‘surround’ so I can control the top-front-side-back distribution more easily (unless you’re a sound engineer, you’ll be surprise how much more this varies than you’d think based on the mix and type of the music).

And, finally, for now, a joystick control to move the balance around between all 11 speakers. Which was done with a single prompt:

> create a joystick control to balance the audio between all the speakers

Next up I’m thinking of keying together listening room profiles per track, keyed into what’s playing on Spotify so the entire balance of the system can auto-configure based on what’s playing. This would then include custom listening profiles and sound stages I’d set up for specific favourite tracks. It’s almost as if one could take all this too far*…?

* Looking out at the hi-fi geek landscape, there is no end of completely stupid nonsense things people do (from £1,000 audiophile USB cables to ‘elevated speaker cable risers‘), none of which make a single bean of difference to anything but your wallet. Once you’ve got a basic system in place by far the biggest differences you’ll be able to actually hear are (a) treating the room with sound insulation, and/or (b) treating the distribution, balance and EQ of the signal itself in your space, which is what I’m doing here.

Working with Dance Collection Danse (DCD) we are initiating a project to create a new national archive: a shared history that will be a co-created and co-managed collaboration between institutions, organizations and individuals.

We believe the idea of an archive in the age of the web isn’t just ‘creating a portal’ but creating an environment where everyone can contribute, access and share.

We are in the design phase, with an aim to  connect records of Canadian dance from across the country into an accessible digital platform to enable engagement, contribution and reuse. The stories of dance and cultural heritage will be made accessible to the general public, as well as artists, educators, researchers, and beyond.

If you are interested in this project, please contact gea@dgen.net.

Light pipes for photons, in triangles and loops

(in the photo, the oblongs are loops of fibre to create delays and the lines all fibre optics)

About 18 months ago (thanks to Libby) I attended a discussion about some of the most advanced quantum computing research in the world.

At the time I thought this must have been how it felt when people first made electrical circuits: the excitement of being able to channel and control electrons, that we now take for granted, but this time with light.

It felt like magic to me that we could control photons within such a beautiful design and wondered, given we see electrical circuits as a visual equivalent to cities, how quantum computers might indicate the future of our city design?

I love Stanza’s work for this (and many) reasons.

“The age of privacy is over. Imagine walking out the door, and knowing every single action, movement, sound, micro movement, pulse, and thread of information is being tracked, monitored, stored, analyzed, interpreted, and logged. The world we will live in seems to be a much bigger brother than the Orwellian vision, it is the mother of big brother.”

What if Nature were not collective but distributive?

Nomos asks if “the Quantum City is, different from the Mechanical and the Dynamic City, not strictly rational” and posits another angle that is highly aligned with contemporary cultural development (from the co-op movement to distributed ledgers/blockchain)

Arida has mapped [2002] the evolution of  world views from the Roman Empire to Quantum theory, and the close links between science  and urban form. He looked at the way we can design and live in quantum cities, exploring methodologies that can frame quantum urbanism: this work continues on the website http://www.quantumcity.com “Urbanism Needs a Conceptual Revolution”.

In one example (of many), IAAC created its framing of the quantised city [video, 2012], blending data from IoT sensors, cloud computing and combining environmental and energy data as part of urban architecture.

CAAD has published [2015] ‘A Quantum City, Mastering the Generic’, self-described as “a love letter to the city and intellectual culture”, it introduces Orlando — neither an authoritarian functionary, nor a restless activist, nor a comfortable member of a bourgeoisie, but a citizen of the digital age, a Quantum Citizen. https://www.caad.arch.ethz.ch/blog/a-quantum-city-mastering-the-generic

It feels to me that there is a line of enquiry to use the subjectivity of art to question the objectivity of science, political dogma, urban design and notions of truth.  

What are the visual loops that could be examined between the circuit designs of quantum computers, city architecture, building design and social constructs?

How does this interface with our notions of Data Infrastructure?

Consolidated Independent (CI) is a world leading digital distribution company.

Acting on behalf of the Independent Music sector, it delivers music, audiobooks and video to iTunes, Amazon, eMusic, Vodafone and literally hundreds of digital retailers.
20% of the music available in the Amazon download store at its launch was delivered by CI.

CI is still the world-leader in pure digital distribution for the sector.

Clients include Beggars Group, Sanctuary, Vital, Mute, etc. One of the most advanced (we built our own cloud compute), detailed, and scalable music-industry platforms in the world.

We enabled albums such as Radiohead’s In Rainbows to go from studio to hundreds of online retailers worldwide within hours of completion.