Wandering mind, engineering student On the AI frontier — models, chips & beyond US stock investor. Theses on the record Connecting with the world 🌍
New York, USA
Joined March 2025
- Tweets336
- Following165
- Followers122
- Likes622
Pinned Tweet
My first open-source project is out: World Pulse 🌍
It's a local-first workspace for following world events. Public RSS reporting, USGS earthquakes, market quotes and AIS ship data all land on one map, and everything runs on your own machine.
It has two modes, and they share your portfolio.
Reading is for keeping up. It gives you a filtered, clustered event stream on the map, organized by region and time, with the health of each source shown. Switch to Portfolio, add your holdings and watchlist, and the feed narrows to reporting that directly mentions those companies.
Command is for watching what moves. It covers the strategic chokepoints: Hormuz, Suez, Bab el-Mandeb and Malacca. You get vessel observations and a timeline you can replay. A side panel shows which of your companies come up in the reports for the area and time window you've selected, and one click takes you back to the full story in Reading.
One design choice I care about: it doesn't guess. A company is only linked when a report mentions it by name or by an alias you've approved. Every link is tagged "verified" or "user-provided", and the "Why connected?" note says plainly that this isn't a price-impact prediction.
There's no account and no hosted backend. The local service listens only on 127.0.0.1, and the cache keeps 7 days of history. AI is optional: you bring your own OpenAI-compatible endpoint, only headlines are sent, and a monthly budget cap stops spending.
Built with TypeScript, Vite and MapLibre. AGPL-3.0, forked from World Monitor.
It's my first real project and some parts are still rough, so feedback is very welcome 👇
github.com/Vesper14916/World…
Claude Duet just launched. And it confirms something I’ve believed about AI for months.
The next AI revolution isn’t about getting better answers.
It’s about not having to ask in the first place.
Anthropic just introduced Claude Duet, a proactive agent that monitors connected apps, works in the background, and reaches out when something needs your attention.
And here’s a detail I really like:
It’s coming to Claude Pro ($20/month) users through a research-preview waitlist.
Meanwhile, OpenAI’s Dots currently requires a Pro or eligible business subscription.
That’s a meaningful step toward making always-on AI more accessible.
But the bigger picture is far more interesting.
OpenAI has Dots. Anthropic has Duet. Meta has Muse. Google is experimenting with Gemini Spark. Grok is moving toward deeper integrations and automation.
Different approaches, but one unmistakable direction:
AI is moving from a tool you use to a responsibility you delegate.
I’ve been experimenting with AI agents and automation for months, and one frustration kept coming back.
I wanted an AI that could monitor my investments 24/7.
But I didn’t want 24/7 notifications.
I didn’t need another chatbot summarizing the same headlines every morning.
I wanted something that remembered my investment thesis, noticed when the underlying facts changed, and interrupted me ONLY when there was a genuine reason to reconsider my position.
That experience made something click.
The real breakthrough isn’t making AI work without prompts. It’s making AI understand when its work actually matters.
Think about the evolution:
2023: “Here’s my question. Give me an answer.”
2025: “Here’s my task. Complete it.”
2026: “Here’s my goal. Keep making progress, and come back when you need me.”
The next step isn’t simply more autonomy.
It’s better judgment.
And I suspect we’re approaching another shift in how we measure AI products.
Today, we obsess over benchmarks, reasoning scores, and model rankings.
Tomorrow, perhaps the metric that matters is:
How much meaningful work gets done per human interruption?
Because an agent that requires constant supervision isn’t really freeing up your time. It’s just creating a new management job.
Of course, there are enormous challenges: trust, privacy, reliability, and knowing when NOT to act.
But I think the direction is becoming clear.
The best AI assistant of the future won’t be the one that talks to you the most.
It’ll be the one that makes you forget how much work it’s doing for you.
And here’s the question I’m genuinely curious about:
If your AI becomes smart enough to decide what deserves your attention, how much control would you actually be willing to give it?
$IREN just put Europe on the map. 🇪🇺
At NVIDIA GTC Berlin, IREN unveiled IREN Europe and a planned ~300 MW AI campus in Badajoz, Spain.
Grid power secured. Civil works license in hand. NVIDIA DSX designs. Target: 2028.
Will this move next quarter’s earnings? No.
Could it change how Wall Street values IREN? Absolutely.
The bigger story isn’t 300 MW. It’s IREN no longer being just a US-focused AI infrastructure play.
Europe wants sovereign AI infrastructure. IREN wants to build it.
Day 10/28.
No improvement. No reset.
Are we testing how long human patience can last now, @thsottiaux? 💀
Schrödinger's portfolio: I'm not losing money until I open the app.(Promise me you won't check your portfolio today. 💀)
Anyone who's worked on a long-running project knows the endless 'Compacting context...' nightmare. 💀
Day 7/
We have increased default context length for GPT-6.1 Sol to 500k when used through Codex & Work, up from 272k or an ~ 80% increase.
We will go higher in the future, but it helps reduce occurrence of compactions and with tasks that require keeping a lot of context and nuance over periods of time. It was carefully tested and is the new sweet spot for most of our users. You can still override it if you want to be higher our lower, with maximum context length of GPT-6.1 Sol of 1M.