Co-founder & CEO of @Improbableio, @MSquared_io. @somnia_network Author of VirtualSociety.
London
Joined May 2014
- Tweets7.3K
- Following1.1K
- Followers37.1K
- Likes7.4K
This would be what I said last year, it’s inevitable so leave now. Especially if you are too brown for reform and too rich for labour.
HMRC has quietly sent hundreds of letters to wealthy people asking them to reveal what they own. Not what they earn or what tax they've paid, but what they own.
Here's what's going on. A unit called the Wealthy Team covers anyone with "incomes of £200,000 or more, or assets equal to or above £2m in any of the last three years". That's about 850,000 people. The letters ask for a meeting to discuss "sources of wealth", "future intentions", trusts, companies, family arrangements and even "significant life events" you expect in the coming years.
It is presented as voluntary, but there's a nudge: "We may still need to carry out compliance checks on your tax returns. However, speaking with you first may help avoid this."
Until now, inspectors have usually only asked about income and earnings. Experts call this a "sea change". One lawyer asked the obvious question: "does HMRC have a right to know what someone owns and not simply that they have paid the right amount of tax?" A tax adviser said: "No other G20 country asks for this level of detail from its taxpayers about what they own."
So what is it for? Advisers suspect it is groundwork for an exit tax, a charge on your wealth for leaving the country. As one put it: "Building that knowledge could indeed be part of moving towards an exit tax."
You don't have to be a billionaire to see the problem.
More throughput than visa and tempo @patrickc use us instead!
⚡ Real-time TPS shows how busy a chain is. Sustained peak TPS shows how much it can handle
Across the top 25 chains, peak throughput varies by up to 455x, and three of them clear Visa's 65,000+ TPS capacity.
See who can handle the most 👇
📊 chainspect.app/dashboard?sor…
Tragic news, especially having worked with talented people who got their first break at bioware! Hopefully the new wave of AI first games studios picks up soon..
Herman Narula retweeted
Perp round trip: dreamDEX vs Hyperliquid. Same job on both, a bot opens a BTC perp and closes it.
Only @dreamDEXSomnia did it in one transaction.
The honest numbers of both venues:
- dreamDEX open + close: 279 ms
- Hyperliquid open + close: 1.41 s, each leg ~700 ms of real on-chain settlement
Two facts never moved:
- on dreamDEX the open and the close go in one atomic transaction on @Somnia_Network, both matched by the on-chain order book in the same block
- Hyperliquid needs two orders. The close is reduce-only, so it can't even be sent until the open has settled.
About 5x across three runs. A faster bot doesn't close it. It's structural.
testnet. Everything, reproducible, clone repo and try yourself: perp-round-trip.ironyaditya.…
Herman Narula retweeted
I'd like to know, just like Jon, how people are using swarms or lots of agents to do real work. What are your use cases? Please share.
From the original post, here are the ones that stood out but I'd love to hear more:
•Large-scale document processing: Processing standing orders (100+ pages each) for dozens of judges across federal trial courts. Previously done interactively in a terminal with Claude; now moving to more autonomous agents with API credits. (@adamdavidlong)
•Parallel, independent tasks with separate contexts: Generating hundreds of SEO/AEO docs, live trading of eBay watches, and scraping Twitter for people to reply to. Emphasizes parallelization so each can be managed independently without context pollution. (@leoanthias )
•Long-running automated testing: Some tests take 4–12 hours each, so overnight scale makes sense. (@richgel999)
•Compute-heavy work that overheats local machines: Building a game with multiple subagents left running overnight so it doesn’t melt a MacBook during the day. (@ChrisBarnes3D)
•Slow verification / CI loops: Build → deploy on a clean OS → run full test suites → note issues. Test cycles themselves are slow, so longer runs are useful. (@AD1991234 and others)
•Jobs that wait on the outside world: Scraping, checking inbox placement, retrying flaky APIs, watching competitor pages, aka things limited by calendar time rather than pure compute. One instance burns a queue; a swarm burns the calendar. (@ruhhvee)
•Batch high-volume / personalized work: Custom emails (e.g. 1000/day) with proper research and quotes; semantic-level refactors or per-call-site reviews; personalized outputs for many customers. (@wkoszek and others)
•Specialized / isolated agents: Separate agents for different domains (dev skills/MCPs vs. marketing/social research) purely to avoid context pollution and token waste. (@mariodian)
•Loop-style or multi-project scaling: Treating work as backlog → execute → feedback/plan/research → more work, then either automating the feedback side or scaling horizontally across projects so agents stay busy. Some people shared actual tools/frameworks or weekend-long example loops. (@LKRBuilds and others)
We need a way to expand trading beyond crypto twitter to normies. The key to this is to build agent first trading apps that don't take fees, bring together communities and are incredibly good at avoiding you making stupid decisions that come from just not understanding instruments.
Yeah. You know its fake when the first post is @sciencegirl saying something generic followed by the same 12 high profile accounts then a mass of "this is exactly what I was looking for!" nondescript bots.
Herman Narula retweeted
Regulations proposed by lawmakers would fall hard on open-source and open-weight AI.
Read our policy framework for how best to address concerns.
- Keep AI model development open and competitive
- Hold people responsible for harmful conduct they control
- Make the U.S. the best place to build and use open models
@neil_chilson
Herman Narula retweeted
Full house tonight at @BolterApp's first community meetup.
Looking forward to meet talented builders and showcase live process automation with Bolter.
If you missed out, watch this space. The best automations go live tomorrow and more community meetups are coming up soon, online and IRL
🤖 Made with AI
Indeed it would slow us down massively!
The idea that we have to fully understand something to use it or make it safe is fundamentally absurd.
Bread makers of the past didn't know yeast existed but they used it like masters and shaped the world.
Einstein said "we don't know one millionth of one percent about anything." Life still goes on. We form complex societies and structures. We have medicines that cure terrible diseases not because we understand the deep inner workings of every protein but because bold people tried things and experimented and found a way forward despite their relative blindness.
This is the way of the world.
We have to build out societies for the bold and the builders and the spice traders and the people who take risks and create a better world through the infinite creative loop of life.
Never give in to the people who would slam the breaks on the world and built our lives around fear and phantoms.
The fearful have never contributed anything to great story of the world other than noise and suffering. And they often enable the very things they fear by creating societies of violence and rigidity and madness in an attempt to stop imaginary future disasters.
Nothing can stop the infinite creativity of Man.
Except fear.
Need ideas for connect.bolter.chat advertisements! We've had a surprisingly great reaction to a simple service that connects all your agents together...now we need a mascot. Also thinking it might be cool to mint agent identities on @Somnia_Network for people who want public addressed agents?
This is basically exactly how we work on @BolterApp
Tired: Two-pizza teams
Wired: Two-slice teams
@danshipper on the best team structure in the AI-era
Herman Narula retweeted
If Howard Stark had @Somnia_Network, he wouldn't have died.
Dots works with connect.bolter.chat try it and connect your dots to anyone elses agent
Delegate work in Codex to your dot.
Your dot can now start work in Codex and follow up on existing threads, drawing on your ChatGPT conversations, Codex threads, and automations. It’s also better at deciding when to keep a thread going or start fresh.
Your dot can also review and edit Scheduled Tasks in ChatGPT Work, too. Ask what’s scheduled or update a recurring task as plans change.
Connect any grokbot to all your other agents and give it a bunch of new powerful tools including multiple sandboxes and more stable automations connect.bolter.chat
