Lead Data Engineer at the University of Glasgow, exploring the future of data and AI through autonomous agents, and agentic workflows
Scotland, United Kingdom
Joined June 2023
- Tweets1.7K
- Following598
- Followers320
- Likes951
Pinned Tweet
NVIDIA Open-Sources Auto Ontology, a Semantic Layer for AI Agents
Your boss asks the AI agent: "What was revenue last quarter?"
It answers in 3 seconds. Clean number. Nice chart.
But it used the wrong "revenue" column. ๐ก๐ผ๐ฏ๐ผ๐ฑ๐ ๐ป๐ผ๐๐ถ๐ฐ๐ฒ๐.
That's the problem NVIDIA's new open-source library, Auto Ontology, goes after. Its README puts it plainly:
"โฆ๐ข ๐ณ๐ฆ๐ท๐ฆ๐ฏ๐ถ๐ฆ ๐ฎ๐ฆ๐ต๐ณ๐ช๐ค ๐ข๐ฏ๐ด๐ธ๐ฆ๐ณ๐ฆ๐ฅ ๐ง๐ข๐ด๐ต๐ฆ๐ณ ๐ช๐ด ๐ธ๐ฐ๐ณ๐ต๐ฉ ๐ฏ๐ฐ๐ต๐ฉ๐ช๐ฏ๐จ ๐ช๐ง ๐ต๐ฉ๐ฆ ๐ฏ๐ถ๐ฎ๐ฃ๐ฆ๐ณ ๐ช๐ด ๐ธ๐ณ๐ฐ๐ฏ๐จ."
๐ช๐ต๐ฎ๐ ๐ถ๐ ๐ถ๐:
A "context layer" that sits between an AI agent and your database. It tells the agent where the data lives, what each field means, and whether to use it.
๐๐ผ๐ ๐ถ๐ ๐๐ผ๐ฟ๐ธ๐, ๐๐๐ฒ๐ฝ ๐ฏ๐ ๐๐๐ฒ๐ฝ:
โข ๐๐ถ๐๐ฐ๐ผ๐๐ฒ๐ฟ: it reads your database metadata and drafts an ontology, a map of business terms like "active customer"
โข ๐๐ฟ๐ผ๐๐ป๐ฑ: agents use those terms instead of guessing from raw table names
โข ๐ฉ๐ฎ๐น๐ถ๐ฑ๐ฎ๐๐ฒ: it checks generated SQL against certified definitions
โข ๐๐
๐ฝ๐ผ๐๐ฒ: agents reach it through MCP, a CLI or an API
โข ๐๐ฒ๐ฎ๐ฟ๐ป: certified corrections go back into the map
๐๐ผ๐ฟ ๐๐๐ฒ๐ฟ๐:
Ask a question, get ๐๐ต๐ฒ ๐ฎ๐ป๐๐๐ฒ๐ฟ, ๐๐ต๐ฒ ๐ฆ๐ค๐ ๐ถ๐ ๐ฟ๐ฎ๐ป, ๐ฎ๐ป๐ฑ ๐๐ต๐ฒ ๐ฟ๐ผ๐๐. You can check the work, not just trust it.
๐๐ผ๐ฟ ๐ฏ๐๐ถ๐น๐ฑ๐ฒ๐ฟ๐:
โข Apache 2.0, runs with Docker Compose or Helm
โข Works with Databricks, Postgres, Snowflake and DuckDB
โข Defaults to NVIDIA's Nemotron models, but ๐ฎ๐ป๐ ๐ข๐ฝ๐ฒ๐ป๐๐-๐ฐ๐ผ๐บ๐ฝ๐ฎ๐๐ถ๐ฏ๐น๐ฒ ๐ฒ๐ป๐ฑ๐ฝ๐ผ๐ถ๐ป๐ ๐๐ผ๐ฟ๐ธ๐
โข The MCP server has ๐ต ๐ฟ๐ฒ๐ฎ๐ฑ-๐ผ๐ป๐น๐ ๐๐ผ๐ผ๐น๐. There is deliberately no tool to browse raw tables, so agents must go through the business terms
๐๐ ๐ฏ๐๐ถ๐น๐ฑ๐ ๐ผ๐ป ๐ฒ๐ฎ๐ฟ๐น๐ถ๐ฒ๐ฟ ๐๐ผ๐ฟ๐ธ:
โข Cube and dbt MetricFlow: semantic layers you write by hand in code
โข Wren AI: an open semantic layer for AI agents, with a context layer built partly by agents
โข Snowflake's Semantic View Autopilot: also drafts semantic models with AI, but only for Snowflake
What's different here: ๐ผ๐ฝ๐ฒ๐ป ๐๐ผ๐๐ฟ๐ฐ๐ฒ, ๐๐ผ๐ฟ๐ธ๐ ๐ฎ๐ฐ๐ฟ๐ผ๐๐ ๐ณ๐ผ๐๐ฟ ๐ฑ๐ฎ๐๐ฎ๐ฏ๐ฎ๐๐ฒ๐, ๐ฎ๐ป๐ฑ ๐ฏ๐๐ถ๐น๐ ๐ณ๐ผ๐ฟ ๐ฎ๐ด๐ฒ๐ป๐๐ ๐๐ต๐ฟ๐ผ๐๐ด๐ต ๐ ๐๐ฃ.
๐ง๐ต๐ฒ ๐ต๐ผ๐ป๐ฒ๐๐ ๐ฝ๐ฎ๐ฟ๐:
โข ๐ง๐ต๐ฒ "๐ต๐ฌ% ๐ณ๐ฎ๐๐๐ฒ๐ฟ" ๐ฐ๐น๐ฎ๐ถ๐บ ๐ถ๐ ๐ก๐ฉ๐๐๐๐'๐ ๐ผ๐๐ป ๐๐ฒ๐๐๐ถ๐ป๐ด, with no baseline published
โข ๐ก๐ผ ๐ฎ๐ฐ๐ฐ๐๐ฟ๐ฎ๐ฐ๐ ๐๐ฐ๐ผ๐ฟ๐ฒ๐ ๐ฎ๐ฟ๐ฒ ๐ฝ๐๐ฏ๐น๐ถ๐๐ต๐ฒ๐ฑ. NVIDIA ships an eval harness (BIRD, 1,534 questions) but no results
โข That harness was tested with ๐๐น๐ฎ๐๐ฑ๐ฒ ๐ข๐ฝ๐๐ ๐ฐ.๐ด, ๐ป๐ผ๐ ๐ก๐ฒ๐บ๐ผ๐๐ฟ๐ผ๐ป
โข ๐ช๐ถ๐๐ต๐ผ๐๐ ๐๐ฎ๐๐ต๐ถ๐๐ผ๐ฟ๐ฝ ๐ฉ๐ฎ๐๐น๐, ๐ฑ๐ฎ๐๐ฎ๐ฏ๐ฎ๐๐ฒ ๐ฝ๐ฎ๐๐๐๐ผ๐ฟ๐ฑ๐ ๐ฎ๐ฟ๐ฒ ๐๐๐ผ๐ฟ๐ฒ๐ฑ ๐๐ป๐ฒ๐ป๐ฐ๐ฟ๐๐ฝ๐๐ฒ๐ฑ in Postgres (per the README)
โข It's young: ๐ฎ๐ฏ๐ผ๐๐ ๐ณ๐ฑ๐ฌ ๐๐ถ๐๐๐๐ฏ ๐๐๐ฎ๐ฟ๐, and NVIDIA isn't accepting outside contributions
๐ง๐ต๐ฒ ๐ฏ๐ถ๐ด๐ด๐ฒ๐ฟ ๐ฝ๐ผ๐ถ๐ป๐: the hard part of AI analytics isn't writing SQL. ๐๐'๐ ๐ฎ๐ด๐ฟ๐ฒ๐ฒ๐ถ๐ป๐ด ๐ผ๐ป ๐๐ต๐ฎ๐ ๐๐ต๐ฒ ๐๐ผ๐ฟ๐ฑ๐ ๐บ๐ฒ๐ฎ๐ป.
GitHub: github.com/NVIDIA/auto-ontolโฆ
Links in reply ๐
๐ง๐ต๐ฒ ๐๐๐น๐น ๐ฆ๐๐ผ๐ฟ๐ ๐ข๐ณ ๐ง๐ต๐ฒ $๐ฒ.๐ฏ๐ ๐๐ถ๐ป๐ฎ๐ป๐ฐ๐ฒ ๐๐ฎ๐ฐ๐ธ, ๐๐ฟ๐ผ๐บ ๐ ๐๐ผ๐ฏ ๐ข๐ณ๐ณ๐ฒ๐ฟ ๐ง๐ผ ๐๐บ๐ฝ๐๐ ๐ช๐ฎ๐น๐น๐ฒ๐๐
The biggest crypto theft in history started with a job interview.
A software engineer at Binance, the worldโs largest crypto exchange, had a profile on a job site. It mentioned they worked at Binance.
One day a message came in from a โProject Directorโ at another company. There was a role. Interested?
First came an interview with someone from HR. Then a second round, with the companyโs CEO.
None of it was real. Not the HR person, not the CEO, not the company. On the other side were North Korean hackers, and they were about to take $6.3 billion.
Hereโs how they did it, based on what Binance has now made public.
After the interviews, the โcompanyโ sent a take-home assignment. Thatโs normal in tech hiring. You download a small task, work on it, send it back.
The engineer downloaded it.
Hidden inside was a way into the laptop. From then on, the hackers could use that engineerโs work computer as if they were sitting at it.
But the laptop was just the front door. What they really wanted was Binanceโs wallets.
In crypto, a wallet is where coins are kept. Whoever controls it can send those coins anywhere, and once theyโre sent, thereโs no bank to call and no โundoโ button. Binance holds billions of dollars of its customersโ coins in its wallets. So the hackers went after the engineers who build and run them.
They already knew who those people were, because a second trap had been running at the same time.
The hackers ran an online ad pushing a fake version of Homebrew, a free tool developers use to install software on their computers. Another Binance engineer clicked the ad and installed the fake one. Binance caught it and removed it. Too late. The hackers had already learned that engineerโs role, and who else at Binance managed the wallets.
So now they had a laptop inside Binance and a list of names.
Next, they turned Binanceโs own tools against it. Binance gives staff an internal AI tool for their work. From the hacked laptop, the attackers used it to take over two more engineers: one who runs the wallet systems day to day, and one who writes wallet software. Binance hasnโt said exactly how.
On one of those computers they found a โcontainer deployment permission token.โ
In plain English: big companies run their software inside sealed boxes called containers. A deployment token is a digital pass that lets you launch or change what runs in those boxes.
Binance says the attacker got into one of these containers, a working part of Binanceโs own system, and from there sent out withdrawals Binance never approved.
Thatโs the part that should scare every company. They never stole the walletsโ private keys, the master passwords to crypto wallets. Binance says thereโs no sign any key leaked. They didnโt need one. They got into Binanceโs own machinery and had it send the money out.
On Thursday, October 29, they took over a series of Binance wallets and drained them.
$6.3 billion. More than four times the old record, the roughly $1.5 billion stolen from Bybit in 2025, also blamed on North Korea.
Binance calls it a โmonthslong scheme.โ That means all of this happened before a single coin moved: a fake company, a fake HR person, a fake CEO doing a second-round interview, a fake download pushed through ads, and quietly mapping out who did what inside Binance.
Some of the money was saved. Part of what was taken was stablecoins, crypto tokens pegged 1:1 to the US dollar. Unlike Bitcoin, theyโre issued by companies, and those companies can freeze them. Tether and Circle froze the stolen ones, and about $2.5 billion has been recovered.
Binance shut withdrawals for about 48 hours. It says it has refilled 98% of the stolen coins from its own reserves and that customer funds are safe.
Two details stay with me.
Binance has a policy that employees in roles that arenโt public-facing shouldnโt have LinkedIn profiles. It had already seen hackers targeting its staff on LinkedIn, and it says it stopped 57 North Korea-linked hacking attempts this year.
This one it didnโt stop.
And none of it needed a genius hack to get started. It needed a believable job offer and one engineer willing to do a take-home test.
Links in First Reply ๐๐
โThis should not have happened,โ Binance said.
thehackernews.com/2026/10/biโฆ
Fortune (the main story: fake recruiter, interviews, Homebrew, AI tool, token, 98% refilled, $2.5B recovered, 57 attempts, LinkedIn policy, Binanceโs quote)
fortune.com/2026/11/03/binanโฆ
NBC News (hacked laptop โ container โ withdrawals Binance never approved; Tetherโs freeze)
nbcnews.com/business/marketsโฆ
Wikipedia, 2026 Binance hack (the date: Thursday, 29 October)
en.wikipedia.org/wiki/2026_Bโฆ
๐๐น๐น ๐ฌ๐ผ๐ ๐ก๐ฒ๐ฒ๐ฑ ๐๐ผ ๐๐ป๐ผ๐ ๐๐ฏ๐ผ๐๐ ๐๐น๐ฎ๐๐ฑ๐ฒ'๐ ๐ก๐ฒ๐ ๐ ๐ฒ๐๐๐ฎ๐ด๐ฒ ๐ง๐ต๐ฟ๐ฒ๐ฎ๐ฑ๐
Picture a coding agent two hours into a job. It has opened 40 files, run the tests a dozen times and written a long plan.
Now it wants to take one tiny next step. Until now, to do that, it had to send Anthropic's API ๐๐ต๐ฒ ๐ฒ๐ป๐๐ถ๐ฟ๐ฒ ๐ฐ๐ผ๐ป๐๐ฒ๐ฟ๐๐ฎ๐๐ถ๐ผ๐ป ๐ฎ๐ด๐ฎ๐ถ๐ป. Every file, every test log, every earlier message. Then again on the next step. And the next.
๐ข๐ป ๐ก๐ผ๐๐ฒ๐บ๐ฏ๐ฒ๐ฟ ๐ฎ, ๐๐ป๐๐ต๐ฟ๐ผ๐ฝ๐ถ๐ฐ ๐ฎ๐ฑ๐ฑ๐ฒ๐ฑ "๐บ๐ฒ๐๐๐ฎ๐ด๐ฒ ๐๐ต๐ฟ๐ฒ๐ฎ๐ฑ๐", ๐ฎ๐ป๐ฑ ๐๐ต๐ฎ๐ ๐ฐ๐ต๐ฎ๐ป๐ด๐ฒ๐.
Here's the difference, with made-up sizes just to show the idea:
โ ๐๐ฒ๐ณ๐ผ๐ฟ๐ฒ: ๐ป๐ผ ๐๐ต๐ฟ๐ฒ๐ฎ๐ฑ
Turn 1: send the whole conversation (800 KB)
Turn 2: send all of it again, plus your new message (802 KB)
Turn 3: all of it again, plus the next message (804 KB)
โ
๐๐ณ๐๐ฒ๐ฟ: ๐๐ถ๐๐ต ๐ฎ ๐๐ต๐ฟ๐ฒ๐ฎ๐ฑ
Turn 1: send the whole conversation once (800 KB)
Turn 2: send only your new message and the ID of Claude's last reply (about 2 KB)
Turn 3: only the next message and the latest ID (about 2 KB)
Anthropic's servers stitch the full conversation back together before Claude reads it. So Claude still sees everything. Your app just stops carrying it back and forth.
Those sizes aren't far from real life. Anthropic says the typical request in long Claude Code sessions went from ๐ฎ๐ฏ๐ผ๐๐ ๐ด๐ฌ๐ฌ ๐๐ ๐๐ผ ๐ญ.๐ด ๐๐.
๐ฆ๐ผ ๐๐ต๐ฎ๐ ๐ต๐ฎ๐ฝ๐ฝ๐ฒ๐ป๐ ๐๐ผ ๐๐ถ๐บ๐ฒ ๐ฎ๐ป๐ฑ ๐ฝ๐ฟ๐ถ๐ฐ๐ฒ?
โฑ๏ธ ๐ง๐ถ๐บ๐ฒ ๐ด๐ผ๐ฒ๐ ๐ฑ๐ผ๐๐ป. Without a thread, the API has to receive, check and tokenize the whole conversation again before Claude can start. With a thread, most of that work covers only the new part. On really big sessions (500K to 1M tokens), Anthropic says Claude started replying ๐ฎ๐ฏ๐ผ๐๐ ๐ฏ๐ฑ% ๐๐ผ๐ผ๐ป๐ฒ๐ฟ.
๐ฐ ๐ฃ๐ฟ๐ถ๐ฐ๐ฒ ๐๐๐ฎ๐๐ ๐๐ต๐ฒ ๐๐ฎ๐บ๐ฒ. Claude still reads the full conversation on every turn, so you pay for the same tokens as before. Rate limits don't change either. If you want to cut cost, that's still the job of prompt caching.
People get confused about who this affects, so let's split it in two.
๐๐ณ ๐๐ผ๐ ๐ท๐๐๐ ๐๐๐ฒ ๐๐น๐ฎ๐๐ฑ๐ฒ ๐๐ผ๐ฑ๐ฒ
You don't have to do anything. Anthropic says threads are already ๐ผ๐ป ๐ฏ๐ ๐ฑ๐ฒ๐ณ๐ฎ๐๐น๐ ๐ณ๐ผ๐ฟ ๐บ๐ผ๐๐ ๐๐น๐ฎ๐๐ฑ๐ฒ ๐๐ผ๐ฑ๐ฒ ๐๐๐ฒ๐ฟ๐.
It all happens in the background. There's no new button, no setting and nothing on screen. You chat and code the way you always have. In long sessions, replies may just start a bit sooner.
If you step away for a while, the thread expires ๐ฑ ๐บ๐ถ๐ป๐๐๐ฒ๐ after the last request. The full history then gets sent again to start a new one. You won't see that happen. That one reply just won't get the speed boost.
Anthropic has only talked about Claude Code here. It hasn't said anything about the regular Claude chat app.
๐๐ณ ๐๐ผ๐ ๐ฏ๐๐ถ๐น๐ฑ ๐ผ๐ป ๐๐ต๐ฒ ๐๐ฃ๐
Here it's the opposite: ๐ป๐ผ๐๐ต๐ถ๐ป๐ด ๐ฐ๐ต๐ฎ๐ป๐ด๐ฒ๐ ๐๐ป๐น๐ฒ๐๐ ๐๐ผ๐ ๐๐๐ฟ๐ป ๐ถ๐ ๐ผ๐ป.
The easiest way is the thread helper in Anthropic's SDK. In Python it looks like this:
client = anthropic.Anthropic()
thread = client.beta.messages.thread(
model="claude-opus-5-5",
max_tokens=1024,
system="You are a coding assistant.")
thread.send({"role": "user", "content": "Fix the failing test."})
thread.send({"role": "user", "content": "Now run it again."})
The first send() sends the full conversation and starts the thread. Every send() after that sends only the new message. The helper adds the beta header for you, and if a thread has expired, it sends the full history again on its own.
A few things worth knowing before you switch it on:
If your chats are only a few thousand tokens long, you probably won't notice much.
This is built for long-running agents.
๐ข๐๐ต๐ฒ๐ฟ๐ ๐ด๐ผ๐ ๐ต๐ฒ๐ฟ๐ฒ ๐ณ๐ถ๐ฟ๐๐, ๐ถ๐ป ๐ฎ ๐ฑ๐ถ๐ณ๐ณ๐ฒ๐ฟ๐ฒ๐ป๐ ๐๐ฎ๐
OpenAI's Responses API and Google's Interactions API can already continue a conversation from an ID.
๐๐ผ๐ฟ ๐น๐ผ๐ป๐ด-๐ฟ๐๐ป๐ป๐ถ๐ป๐ด ๐ฎ๐ด๐ฒ๐ป๐๐, ๐ท๐๐๐ ๐บ๐ผ๐๐ถ๐ป๐ด ๐๐ต๐ฒ ๐ฐ๐ผ๐ป๐๐ฒ๐ฟ๐๐ฎ๐๐ถ๐ผ๐ป ๐ฎ๐ฟ๐ผ๐๐ป๐ฑ ๐ต๐ฎ๐ ๐ฏ๐ฒ๐ฐ๐ผ๐บ๐ฒ ๐ฝ๐ฎ๐ฟ๐ ๐ผ๐ณ ๐๐ต๐ฒ ๐๐ฎ๐ถ๐.
claude.com/resources/articleโฆ
Gul Saeed Khattak retweeted
Saw this and wondered why Elon still works like this when he doesn't need to. I got the answer pretty quickly.
Tbh a yacht is fun for a week.
After that it's the same sea every morning.
Some might disagree , but that is fine .
Most of us think the point of money is to stop working. But rest only feels good when there's work around it. Take the work away and relaxing becomes its own kind of boredom.
What doesn't get boring is building something nobody has built before. You never really finish it, so it keeps you hooked for years.
If Elon had cashed out and moved onto a boat, he wouldn't be Elon, and none of us would be talking about his 3am office nights.
People with lots of money and nothing to build often drift into bad habits. People deep in building something rarely have time for them.
On a much smaller scale, my best late nights are the ones spent building something.
It feels more like rest than any day off.
An employee who worked closely with Elon Musk for four years revealed that he had seen Elon working in the office countless times at 1, 2, even 3 a.m.
Elon works every waking hour. He doesnโt go out for lunch or dinner. He mostly has his food brought into meetings and eats while the meeting continues.
As the richest person on the planet, he could be anywhere in the world, relaxing on a yacht and enjoying his wealth. Instead, he keeps showing up at the office, working late, and giving his precious time to building a better future for humanity.
Huge respect for Elon. The dedication he puts into building a better future deserve far more recognition.
Gul Saeed Khattak retweeted
๐ก๐๐ถ๐ฑ๐ถ๐ฎ'๐ ๐ฑ๐ญ% ๐ฅ๐๐น๐ฒ: ๐ช๐ต๐ฎ๐ ๐๐ ๐๐ถ๐๐ฒ๐ ๐ฌ๐ผ๐ ๐๐ป๐ฑ ๐ช๐ต๐ฎ๐ ๐๐ ๐๐ผ๐ฒ๐๐ป'๐
Picture a data centre in Singapore. Racks full of Nvidia chips. Its biggest clients are Chinese AI companies that can't buy those chips at home, so they rent them.
๐ง๐ต๐ฒ ๐จ๐ป๐ฑ๐ฒ๐ฟ๐น๐๐ถ๐ป๐ด ๐๐ฑ๐ฒ๐ฎ
Compute is a physical thing, but access to it isn't. You can stop a box at a border. It's much harder to stop someone logging in to a box in another country.
So there are two doors:
โข ๐๐๐๐ถ๐ป๐ด: export controls, plus Nvidia's "white list", which per the FT more than halved approved Asian buyers this year.
โข ๐ฅ๐ฒ๐ป๐๐ถ๐ป๐ด: Chinese companies leasing GPUs in overseas data centres, including for model training.
Close the first door and traffic moves to the second.
๐ง๐ต๐ถ๐ ๐ฟ๐๐น๐ฒ ๐ถ๐ ๐ฎ ๐๐ฎ๐น๐๐ฒ ๐ผ๐ป ๐๐ต๐ฒ ๐๐ฒ๐ฐ๐ผ๐ป๐ฑ ๐ฑ๐ผ๐ผ๐ฟ.
๐๐ป ๐๐ป๐ฎ๐น๐ผ๐ด๐
It's a bit like a landlord telling a building manager: no more than 49% of tenants from one place. Not a law. A lease condition, enforced by whoever signs the lease.
๐ก๐๐ถ๐ฑ๐ถ๐ฎ ๐ป๐ผ๐ ๐๐ฎ๐ป๐๐ ๐๐ต๐ฎ๐ ๐๐ผ ๐ฐ๐ต๐ฎ๐ป๐ด๐ฒ.
๐ช๐ต๐ฎ๐ ๐๐ฎ๐ฝ๐ฝ๐ฒ๐ป๐ฒ๐ฑ
According to the Financial Times (five people familiar with the matter), Nvidia has told cloud providers that ๐ฎ๐ ๐น๐ฒ๐ฎ๐๐ ๐ฑ๐ญ% of their services must go to non-Chinese companies. There was no such limit before.
โข ๐ก๐ผ ๐ฝ๐ฎ๐ฝ๐ฒ๐ฟ๐๐ผ๐ฟ๐ธ: the instruction was given verbally, through Nvidia's sales and regional teams.
โข ๐ช๐ผ๐ฟ๐น๐ฑ๐๐ถ๐ฑ๐ฒ: it applies to clients globally, not only in Asia (two people told the FT).
โข ๐ก๐ฒ๐ ๐ฑ๐ฒ๐ฎ๐น๐ ๐ผ๐ป๐น๐: existing leases are not affected.
โข ๐ฆ๐ถ๐ป๐ฐ๐ฒ ๐๐ต๐ฒ ๐๐๐บ๐บ๐ฒ๐ฟ: two people said Nvidia did it pre-emptively, to cut criticism from US lawmakers.
๐ช๐ต๐ ๐ฅ๐ฒ๐ป๐๐ถ๐ป๐ด ๐ ๐ฎ๐๐๐ฒ๐ฟ๐
US export rules mostly control where chips are shipped. Renting them in a data centre abroad has been the gap. Many Chinese groups used overseas leases for heavy AI work, including model training.
Nvidia had already tightened the front door. This year it built a "white list" of approved Asian buyers and ๐บ๐ผ๐ฟ๐ฒ ๐๐ต๐ฎ๐ป ๐ต๐ฎ๐น๐๐ฒ๐ฑ it, per the FT. That pushed more Chinese firms toward renting.
This new rule goes after ๐๐ต๐ฒ ๐ฟ๐ฒ๐ป๐๐ถ๐ป๐ด.
๐ช๐ต๐ฎ๐ ๐๐ ๐๐ผ๐ฒ๐ ๐ง๐ผ ๐ฃ๐ฟ๐ถ๐ฐ๐ฒ๐
A rack of B300 chips sells for ๐ฎ๐ฏ๐ผ๐๐ $๐ฒ๐ฌ๐ฌ,๐ฌ๐ฌ๐ฌ internationally. One person told the FT a rack fetched ๐ฅ๐บ๐ฏ๐ญ๐ด.๐ฑ๐บ๐ป (๐ฎ๐ฏ๐ผ๐๐ $๐ฎ.๐ด๐บ๐ป) in China last month.
That is ๐ฟ๐ผ๐๐ด๐ต๐น๐ ๐ฐ.๐ณ ๐๐ถ๐บ๐ฒ๐ the price. The B300 can't legally be sold in China.
๐๐ผ๐ฟ ๐๐น๐ผ๐๐ฑ ๐ฃ๐ฟ๐ผ๐๐ถ๐ฑ๐ฒ๐ฟ๐
Chinese tech groups are the biggest clients for data centres in places like south-east Asia. A Singapore-based provider told the FT it is hard to find enough non-Chinese clients outside the US.
Providers that can't find them may have to ๐๐ต๐ฟ๐ถ๐ป๐ธ.
New contracts only.
๐ง๐ต๐ฒ ๐ ๐ฎ๐ฝ ๐ข๐ณ ๐๐ต๐ถ๐ป๐ฎ ๐๐ผ๐บ๐ฝ๐๐๐ฒ ๐๐ผ๐ป๐๐ฟ๐ผ๐น๐, ๐ข๐ป๐ฒ ๐๐ถ๐ป๐ฒ ๐๐ฎ๐ฐ๐ต
โข ๐จ๐ฆ ๐ฒ๐
๐ฝ๐ผ๐ฟ๐ ๐ฐ๐ผ๐ป๐๐ฟ๐ผ๐น๐: block shipping top chips like the B300 to China.
โข ๐๐๐ฆ ๐ฟ๐๐น๐ฒ (๐๐ฎ๐ป๐๐ฎ๐ฟ๐): license applicants must list remote cloud users in China and other listed countries.
โข ๐๐๐ฆ ๐ด๐๐ถ๐ฑ๐ฎ๐ป๐ฐ๐ฒ (๐ ๐ฎ๐ ๐ฏ๐ญ): a license is needed for advanced chips going to companies headquartered in China or Macau, wherever they are.
โข ๐ก๐๐ถ๐ฑ๐ถ๐ฎ'๐ ๐๐ต๐ถ๐๐ฒ ๐น๐ถ๐๐ (๐๐ต๐ถ๐ ๐๐ฒ๐ฎ๐ฟ): approved Asian buyers more than halved, per the FT.
โข ๐ฅ๐ฒ๐บ๐ผ๐๐ฒ ๐๐ฐ๐ฐ๐ฒ๐๐ ๐ฆ๐ฒ๐ฐ๐๐ฟ๐ถ๐๐ ๐๐ฐ๐: passed the House 369-22 in January; sitting with the Senate Banking Committee.
โข ๐ก๐๐ถ๐ฑ๐ถ๐ฎ'๐ ๐ฑ๐ญ% ๐ฟ๐๐น๐ฒ (๐ป๐ผ๐): aimed at the cloud provider's customer mix, not the chip.
๐ช๐ต๐ฒ๐ฟ๐ฒ ๐ง๐ต๐ถ๐ ๐๐ถ๐๐
Government rules set the floor. Nvidia is adding its own layer on top, so lawmakers have less to criticise (two people told the FT that was the motive).
๐ช๐ต๐ผ ๐๐ฒ๐ฒ๐น๐ ๐๐
โข ๐๐น๐ผ๐๐ฑ ๐ฝ๐ฟ๐ผ๐๐ถ๐ฑ๐ฒ๐ฟ๐ ๐ถ๐ป ๐๐ผ๐๐๐ต-๐ฒ๐ฎ๐๐ ๐๐๐ถ๐ฎ: Chinese tech groups are their biggest clients.
โข ๐๐ต๐ถ๐ป๐ฒ๐๐ฒ ๐๐ ๐ฐ๐ผ๐บ๐ฝ๐ฎ๐ป๐ถ๐ฒ๐: already paying a premium. A B300 rack, ๐ฎ๐ฏ๐ผ๐๐ $๐ฒ๐ฌ๐ฌ๐ธ abroad, reportedly fetched $๐ฎ.๐ด๐บ๐ป in China last month (one deal).
๐ง๐ต๐ฒ ๐ฏ๐ถ๐ด๐ด๐ฒ๐ฟ ๐ฝ๐ผ๐ถ๐ป๐: ๐ต๐ฉ๐ฆ ๐ค๐ฉ๐ช๐ฑ ๐ธ๐ข๐ณ ๐ช๐ด ๐ฎ๐ฐ๐ท๐ช๐ฏ๐จ ๐ง๐ณ๐ฐ๐ฎ ๐ธ๐ฉ๐ฐ ๐ค๐ข๐ฏ ๐ฃ๐ถ๐บ ๐ค๐ฉ๐ช๐ฑ๐ด ๐ต๐ฐ ๐ธ๐ฉ๐ฐ ๐ค๐ข๐ฏ ๐ณ๐ฆ๐ฏ๐ต ๐ต๐ฉ๐ฆ๐ฎ.
FT (original report): ft.com/content/1902df70-1380โฆ
Links in reply ๐
Gul Saeed Khattak retweeted
Who is this reply even for?
Saw a reply today under my NVIDIA auto-ontology repo post today .
It just said "@bot save this".
I assumed it was a joke, so I looked it up. @bot is xAI's Grok Bot, an agent app they launched in August 2026.
You tag it under a post and the post goes to your own agent, which runs on its own computer in the cloud all day.
So that reply wasn't really for the person who shared the repo.
It was for someone's private assistant, left out in public where the rest of us could read it.
Feels a bit like leaving a note for your cleaner on a stranger's front door.
I get why people do it though. I have hundreds of bookmarks I've never opened.
Every time I save something I tell myself I'll read it later, and I don't. At least an agent might actually go through the list.
I read that repo properly because I'm building an ontology layer on top of a lineage graph at work. I doubt most people saving it will open the README. Their bots probably will.
Still can't decide how I feel about that.
๐ก๐๐ถ๐ฑ๐ถ๐ฎ'๐ ๐ฑ๐ญ% ๐ฅ๐๐น๐ฒ: ๐ช๐ต๐ฎ๐ ๐๐ ๐๐ถ๐๐ฒ๐ ๐ฌ๐ผ๐ ๐๐ป๐ฑ ๐ช๐ต๐ฎ๐ ๐๐ ๐๐ผ๐ฒ๐๐ป'๐
Picture a data centre in Singapore. Racks full of Nvidia chips. Its biggest clients are Chinese AI companies that can't buy those chips at home, so they rent them.
๐ง๐ต๐ฒ ๐จ๐ป๐ฑ๐ฒ๐ฟ๐น๐๐ถ๐ป๐ด ๐๐ฑ๐ฒ๐ฎ
Compute is a physical thing, but access to it isn't. You can stop a box at a border. It's much harder to stop someone logging in to a box in another country.
So there are two doors:
โข ๐๐๐๐ถ๐ป๐ด: export controls, plus Nvidia's "white list", which per the FT more than halved approved Asian buyers this year.
โข ๐ฅ๐ฒ๐ป๐๐ถ๐ป๐ด: Chinese companies leasing GPUs in overseas data centres, including for model training.
Close the first door and traffic moves to the second.
๐ง๐ต๐ถ๐ ๐ฟ๐๐น๐ฒ ๐ถ๐ ๐ฎ ๐๐ฎ๐น๐๐ฒ ๐ผ๐ป ๐๐ต๐ฒ ๐๐ฒ๐ฐ๐ผ๐ป๐ฑ ๐ฑ๐ผ๐ผ๐ฟ.
๐๐ป ๐๐ป๐ฎ๐น๐ผ๐ด๐
It's a bit like a landlord telling a building manager: no more than 49% of tenants from one place. Not a law. A lease condition, enforced by whoever signs the lease.
๐ก๐๐ถ๐ฑ๐ถ๐ฎ ๐ป๐ผ๐ ๐๐ฎ๐ป๐๐ ๐๐ต๐ฎ๐ ๐๐ผ ๐ฐ๐ต๐ฎ๐ป๐ด๐ฒ.
๐ช๐ต๐ฎ๐ ๐๐ฎ๐ฝ๐ฝ๐ฒ๐ป๐ฒ๐ฑ
According to the Financial Times (five people familiar with the matter), Nvidia has told cloud providers that ๐ฎ๐ ๐น๐ฒ๐ฎ๐๐ ๐ฑ๐ญ% of their services must go to non-Chinese companies. There was no such limit before.
โข ๐ก๐ผ ๐ฝ๐ฎ๐ฝ๐ฒ๐ฟ๐๐ผ๐ฟ๐ธ: the instruction was given verbally, through Nvidia's sales and regional teams.
โข ๐ช๐ผ๐ฟ๐น๐ฑ๐๐ถ๐ฑ๐ฒ: it applies to clients globally, not only in Asia (two people told the FT).
โข ๐ก๐ฒ๐ ๐ฑ๐ฒ๐ฎ๐น๐ ๐ผ๐ป๐น๐: existing leases are not affected.
โข ๐ฆ๐ถ๐ป๐ฐ๐ฒ ๐๐ต๐ฒ ๐๐๐บ๐บ๐ฒ๐ฟ: two people said Nvidia did it pre-emptively, to cut criticism from US lawmakers.
๐ช๐ต๐ ๐ฅ๐ฒ๐ป๐๐ถ๐ป๐ด ๐ ๐ฎ๐๐๐ฒ๐ฟ๐
US export rules mostly control where chips are shipped. Renting them in a data centre abroad has been the gap. Many Chinese groups used overseas leases for heavy AI work, including model training.
Nvidia had already tightened the front door. This year it built a "white list" of approved Asian buyers and ๐บ๐ผ๐ฟ๐ฒ ๐๐ต๐ฎ๐ป ๐ต๐ฎ๐น๐๐ฒ๐ฑ it, per the FT. That pushed more Chinese firms toward renting.
This new rule goes after ๐๐ต๐ฒ ๐ฟ๐ฒ๐ป๐๐ถ๐ป๐ด.
๐ช๐ต๐ฎ๐ ๐๐ ๐๐ผ๐ฒ๐ ๐ง๐ผ ๐ฃ๐ฟ๐ถ๐ฐ๐ฒ๐
A rack of B300 chips sells for ๐ฎ๐ฏ๐ผ๐๐ $๐ฒ๐ฌ๐ฌ,๐ฌ๐ฌ๐ฌ internationally. One person told the FT a rack fetched ๐ฅ๐บ๐ฏ๐ญ๐ด.๐ฑ๐บ๐ป (๐ฎ๐ฏ๐ผ๐๐ $๐ฎ.๐ด๐บ๐ป) in China last month.
That is ๐ฟ๐ผ๐๐ด๐ต๐น๐ ๐ฐ.๐ณ ๐๐ถ๐บ๐ฒ๐ the price. The B300 can't legally be sold in China.
๐๐ผ๐ฟ ๐๐น๐ผ๐๐ฑ ๐ฃ๐ฟ๐ผ๐๐ถ๐ฑ๐ฒ๐ฟ๐
Chinese tech groups are the biggest clients for data centres in places like south-east Asia. A Singapore-based provider told the FT it is hard to find enough non-Chinese clients outside the US.
Providers that can't find them may have to ๐๐ต๐ฟ๐ถ๐ป๐ธ.
New contracts only.
๐ง๐ต๐ฒ ๐ ๐ฎ๐ฝ ๐ข๐ณ ๐๐ต๐ถ๐ป๐ฎ ๐๐ผ๐บ๐ฝ๐๐๐ฒ ๐๐ผ๐ป๐๐ฟ๐ผ๐น๐, ๐ข๐ป๐ฒ ๐๐ถ๐ป๐ฒ ๐๐ฎ๐ฐ๐ต
โข ๐จ๐ฆ ๐ฒ๐
๐ฝ๐ผ๐ฟ๐ ๐ฐ๐ผ๐ป๐๐ฟ๐ผ๐น๐: block shipping top chips like the B300 to China.
โข ๐๐๐ฆ ๐ฟ๐๐น๐ฒ (๐๐ฎ๐ป๐๐ฎ๐ฟ๐): license applicants must list remote cloud users in China and other listed countries.
โข ๐๐๐ฆ ๐ด๐๐ถ๐ฑ๐ฎ๐ป๐ฐ๐ฒ (๐ ๐ฎ๐ ๐ฏ๐ญ): a license is needed for advanced chips going to companies headquartered in China or Macau, wherever they are.
โข ๐ก๐๐ถ๐ฑ๐ถ๐ฎ'๐ ๐๐ต๐ถ๐๐ฒ ๐น๐ถ๐๐ (๐๐ต๐ถ๐ ๐๐ฒ๐ฎ๐ฟ): approved Asian buyers more than halved, per the FT.
โข ๐ฅ๐ฒ๐บ๐ผ๐๐ฒ ๐๐ฐ๐ฐ๐ฒ๐๐ ๐ฆ๐ฒ๐ฐ๐๐ฟ๐ถ๐๐ ๐๐ฐ๐: passed the House 369-22 in January; sitting with the Senate Banking Committee.
โข ๐ก๐๐ถ๐ฑ๐ถ๐ฎ'๐ ๐ฑ๐ญ% ๐ฟ๐๐น๐ฒ (๐ป๐ผ๐): aimed at the cloud provider's customer mix, not the chip.
๐ช๐ต๐ฒ๐ฟ๐ฒ ๐ง๐ต๐ถ๐ ๐๐ถ๐๐
Government rules set the floor. Nvidia is adding its own layer on top, so lawmakers have less to criticise (two people told the FT that was the motive).
๐ช๐ต๐ผ ๐๐ฒ๐ฒ๐น๐ ๐๐
โข ๐๐น๐ผ๐๐ฑ ๐ฝ๐ฟ๐ผ๐๐ถ๐ฑ๐ฒ๐ฟ๐ ๐ถ๐ป ๐๐ผ๐๐๐ต-๐ฒ๐ฎ๐๐ ๐๐๐ถ๐ฎ: Chinese tech groups are their biggest clients.
โข ๐๐ต๐ถ๐ป๐ฒ๐๐ฒ ๐๐ ๐ฐ๐ผ๐บ๐ฝ๐ฎ๐ป๐ถ๐ฒ๐: already paying a premium. A B300 rack, ๐ฎ๐ฏ๐ผ๐๐ $๐ฒ๐ฌ๐ฌ๐ธ abroad, reportedly fetched $๐ฎ.๐ด๐บ๐ป in China last month (one deal).
๐ง๐ต๐ฒ ๐ฏ๐ถ๐ด๐ด๐ฒ๐ฟ ๐ฝ๐ผ๐ถ๐ป๐: ๐ต๐ฉ๐ฆ ๐ค๐ฉ๐ช๐ฑ ๐ธ๐ข๐ณ ๐ช๐ด ๐ฎ๐ฐ๐ท๐ช๐ฏ๐จ ๐ง๐ณ๐ฐ๐ฎ ๐ธ๐ฉ๐ฐ ๐ค๐ข๐ฏ ๐ฃ๐ถ๐บ ๐ค๐ฉ๐ช๐ฑ๐ด ๐ต๐ฐ ๐ธ๐ฉ๐ฐ ๐ค๐ข๐ฏ ๐ณ๐ฆ๐ฏ๐ต ๐ต๐ฉ๐ฆ๐ฎ.
FT (original report): ft.com/content/1902df70-1380โฆ
Links in reply ๐
The Information's "not mandatory" report, summarised: the-decoder.com/nvidia-pusheโฆ
Nvidia's halved Asian white list (July): thenextweb.com/news/nvidia-hโฆ
Remote Access Security Act passes House: mondaq.com/unitedstates/expoโฆ
Saw this and wondered why Elon still works like this when he doesn't need to. I got the answer pretty quickly.
Tbh a yacht is fun for a week.
After that it's the same sea every morning.
Some might disagree , but that is fine .
Most of us think the point of money is to stop working. But rest only feels good when there's work around it. Take the work away and relaxing becomes its own kind of boredom.
What doesn't get boring is building something nobody has built before. You never really finish it, so it keeps you hooked for years.
If Elon had cashed out and moved onto a boat, he wouldn't be Elon, and none of us would be talking about his 3am office nights.
People with lots of money and nothing to build often drift into bad habits. People deep in building something rarely have time for them.
On a much smaller scale, my best late nights are the ones spent building something.
It feels more like rest than any day off.
An employee who worked closely with Elon Musk for four years revealed that he had seen Elon working in the office countless times at 1, 2, even 3 a.m.
Elon works every waking hour. He doesnโt go out for lunch or dinner. He mostly has his food brought into meetings and eats while the meeting continues.
As the richest person on the planet, he could be anywhere in the world, relaxing on a yacht and enjoying his wealth. Instead, he keeps showing up at the office, working late, and giving his precious time to building a better future for humanity.
Huge respect for Elon. The dedication he puts into building a better future deserve far more recognition.
Who is this reply even for?
Saw a reply today under my NVIDIA auto-ontology repo post today .
It just said "@bot save this".
I assumed it was a joke, so I looked it up. @bot is xAI's Grok Bot, an agent app they launched in August 2026.
You tag it under a post and the post goes to your own agent, which runs on its own computer in the cloud all day.
So that reply wasn't really for the person who shared the repo.
It was for someone's private assistant, left out in public where the rest of us could read it.
Feels a bit like leaving a note for your cleaner on a stranger's front door.
I get why people do it though. I have hundreds of bookmarks I've never opened.
Every time I save something I tell myself I'll read it later, and I don't. At least an agent might actually go through the list.
I read that repo properly because I'm building an ontology layer on top of a lineage graph at work. I doubt most people saving it will open the README. Their bots probably will.
Still can't decide how I feel about that.
Gul Saeed Khattak retweeted
Claude Code says no to game clones. A tool that uses Claude Code tops trending anyway.
What changed:
Last week Claude Code got two new sentences telling it not to recreate specific existing games.
What it is:
This week a tool called Craft Anything is topping the trending charts. It reverse engineers apps and games, then has a coding agent build your own version.
1.6K stars today. 5.7K total.
To be fair, it defaults to a clean-room mode and says it never bypasses DRM, logins or anti-cheat.
The funny part:
The builder it supports first is Claude Code. One of its own example requests is "Recreate this game's gameplay in Rust".
My take:
Big labs can block this kind of use inside their own agents. I don't think it changes much. Craft Anything already has adapters for open agents like Qwen Code and opencode. Open models are good enough now, and anyone can point these tools at them.
Reverse engineering always cut both ways. The debugger that fixes your crash can also copy someone's code. What changed is the effort. It used to take a skilled person weeks. Now it's closer to a prompt.
Links in reply ๐
Reply comment (post this straight after):
Repo: github.com/storytold/craft-aโฆ
Gul Saeed Khattak retweeted
OpenAI's $172K bug bounty, and why the partner matters more
My read on OpenAI's security news is a bit different from the headlines.
The money:
โข $172,000 is the top payout, with bonuses stacked onโจโข A critical bug in ChatGPT, Codex or the API now pays up to $30,000โจโข OpenAI says that's more than four times what critical findings paid before
The partner:
The more interesting part is who they picked. Cathedral is the startup whose AI agents reached Anthropic's internal business systems in October 2026.
Not through Claude.
Through third-party business software.
Anthropic disputes parts of it. It says it took several days, not 12 hours, and that Cathedral went beyond the bounty rules by targeting vendors and not disclosing everything.
Now OpenAI brings them in as a partner, not a bounty hunter. I think that's smart. A bounty comes with a fixed scope and safe harbour. A partner can agree up front what's in bounds.
The Anthropic fight was mostly about scope and disclosure.
What I'd watch:
The path that worked at Anthropic was vendor software, not the model. Bigger bounties for ChatGPT bugs don't cover that.
OpenAI hasn't said what Cathedral will test.
Maybe vendors are already in scope.
If not, the money is pointed at the wrong door.
One line from their write-up on the last sprint stuck with me.
A merged patch didn't always mean the fix had reached every deployed system.
I've seen the same in data pipelines. Code merged, old job still running somewhere.
Links in reply ๐
OpenAI's sprint post: openai.com/index/preparing-oโฆ
Gul Saeed Khattak retweeted
๐ฏ ๐ข๐ฝ๐ฒ๐ป-๐ฆ๐ผ๐๐ฟ๐ฐ๐ฒ ๐ง๐ผ๐ผ๐น๐ ๐ฅ๐ฎ๐ฐ๐ถ๐ป๐ด ๐๐ผ ๐ ๐ฎ๐ป๐ฎ๐ด๐ฒ ๐ฌ๐ผ๐๐ฟ ๐๐ ๐๐ด๐ฒ๐ป๐๐
You have Claude Code in one terminal. Codex in another. Five more agents running somewhere.
๐ช๐ต๐ผ'๐ ๐บ๐ฎ๐ป๐ฎ๐ด๐ถ๐ป๐ด ๐๐ต๐ฒ๐บ?
Three open-source projects want that job. Same goal, very different ideas.
๐ฃ๐ฎ๐ฝ๐ฒ๐ฟ๐ฐ๐น๐ถ๐ฝ: ๐๐ต๐ฒ ๐ฐ๐ผ๐บ๐ฝ๐ฎ๐ป๐ (๐ญ๐ญ๐ฌ๐ธ โญ)
"๐๐ง ๐๐ฑ๐ฆ๐ฏ๐๐ญ๐ข๐ธ ๐ช๐ด ๐ข๐ฏ ๐ฆ๐ฎ๐ฑ๐ญ๐ฐ๐บ๐ฆ๐ฆ, ๐๐ข๐ฑ๐ฆ๐ณ๐ค๐ญ๐ช๐ฑ ๐ช๐ด ๐ต๐ฉ๐ฆ ๐ค๐ฐ๐ฎ๐ฑ๐ข๐ฏ๐บ."
โข Hire agents into roles: CEO, engineer, marketer
โข Set goals and budgets, approve plans
โข Not just code: research, content, operations too
๐ข๐ฝ๐ฒ๐ป๐๐ฎ๐ป๐ฑ๐: ๐๐ต๐ฒ ๐ฒ๐ป๐ด๐ถ๐ป๐ฒ๐ฒ๐ฟ๐ถ๐ป๐ด ๐๐ฒ๐ฎ๐บ (๐ต๐ฎ๐ธ โญ)
Now leads with Agent Canvas, a "๐ฅ๐ฆ๐ท๐ฆ๐ญ๐ฐ๐ฑ๐ฆ๐ณ ๐ค๐ฐ๐ฏ๐ต๐ณ๐ฐ๐ญ ๐ค๐ฆ๐ฏ๐ต๐ฆ๐ณ ๐ง๐ฐ๐ณ ๐ค๐ฐ๐ฅ๐ช๐ฏ๐จ ๐ข๐จ๐ฆ๐ฏ๐ต๐ด."
โข The only one of the three with ๐ถ๐๐ ๐ผ๐๐ป ๐ฐ๐ผ๐ฑ๐ถ๐ป๐ด ๐ฎ๐ด๐ฒ๐ป๐
โข Also runs Claude Code, Codex, Gemini and others
โข Automations with Slack, GitHub and Linear. Self-host or use their cloud
๐ฅ๐๐ณ๐น๐ผ: ๐๐ต๐ฒ ๐ฝ๐ผ๐๐ฒ๐ฟ ๐๐ผ๐ผ๐น (๐ด๐ญ๐ธ โญ)
"๐๐ฏ ๐ข๐จ๐ฆ๐ฏ๐ต ๐ฎ๐ฆ๐ต๐ข-๐ฉ๐ข๐ณ๐ฏ๐ฆ๐ด๐ด ๐ง๐ฐ๐ณ ๐๐ญ๐ข๐ถ๐ฅ๐ฆ ๐๐ฐ๐ฅ๐ฆ ๐ข๐ฏ๐ฅ ๐๐ฐ๐ฅ๐ฆ๐น."
โข Lives inside Claude Code as plugins
โข Runs "swarms" of agents, with routing and policy rules
โข Built for developers who stay in the terminal
๐ง๐ต๐ฒ ๐๐ถ๐บ๐ฝ๐น๐ฒ ๐๐ฒ๐ฟ๐๐ถ๐ผ๐ป:
โข Paperclip manages ๐๐ต๐ฒ ๐ฐ๐ผ๐บ๐ฝ๐ฎ๐ป๐
โข OpenHands runs ๐๐ต๐ฒ ๐ฒ๐ป๐ด๐ถ๐ป๐ฒ๐ฒ๐ฟ๐ถ๐ป๐ด ๐๐ฒ๐ฎ๐บ
โข Ruflo powers up ๐๐ต๐ฒ ๐ฑ๐ฒ๐๐ฒ๐น๐ผ๐ฝ๐ฒ๐ฟ
๐ง๐ต๐ฒ ๐๐ถ๐น๐ฑ ๐ฝ๐ฎ๐ฟ๐: Paperclip only launched in ๐ ๐ฎ๐ฟ๐ฐ๐ต ๐ฎ๐ฌ๐ฎ๐ฒ. It passed both in about 8 months.
๐๐ฒ๐ฒ๐ฝ ๐ถ๐ป ๐บ๐ถ๐ป๐ฑ:
โข ๐ฆ๐๐ฎ๐ฟ๐ ๐บ๐ฒ๐ฎ๐๐๐ฟ๐ฒ ๐ฎ๐๐๐ฒ๐ป๐๐ถ๐ผ๐ป, ๐ป๐ผ๐ ๐ฟ๐ฒ๐ฎ๐น ๐๐๐ฒ
โข They overlap but aren't the same product. Pick by how you work, not by stars
The race isn't to build the best agent anymore. ๐๐'๐ ๐๐ผ ๐ฏ๐ฒ ๐๐ต๐ฒ ๐๐ฐ๐ฟ๐ฒ๐ฒ๐ป ๐๐ผ๐ ๐บ๐ฎ๐ป๐ฎ๐ด๐ฒ ๐๐ต๐ฒ๐บ ๐ณ๐ฟ๐ผ๐บ.
๐๐ต๐ข๐ณ ๐ค๐ฐ๐ถ๐ฏ๐ต๐ด ๐ข๐ด ๐ฐ๐ง ๐๐ฐ๐ท 3, 2026.
Links in reply ๐
Claude Code says no to game clones. A tool that uses Claude Code tops trending anyway.
What changed:
Last week Claude Code got two new sentences telling it not to recreate specific existing games.
What it is:
This week a tool called Craft Anything is topping the trending charts. It reverse engineers apps and games, then has a coding agent build your own version.
1.6K stars today. 5.7K total.
To be fair, it defaults to a clean-room mode and says it never bypasses DRM, logins or anti-cheat.
The funny part:
The builder it supports first is Claude Code. One of its own example requests is "Recreate this game's gameplay in Rust".
My take:
Big labs can block this kind of use inside their own agents. I don't think it changes much. Craft Anything already has adapters for open agents like Qwen Code and opencode. Open models are good enough now, and anyone can point these tools at them.
Reverse engineering always cut both ways. The debugger that fixes your crash can also copy someone's code. What changed is the effort. It used to take a skilled person weeks. Now it's closer to a prompt.
Links in reply ๐
Reply comment (post this straight after):
Repo: github.com/storytold/craft-aโฆ
Claude Code's new game rule, reported by Tom's Hardware: tomshardware.com/tech-industโฆ
OpenAI's $172K bug bounty, and why the partner matters more
My read on OpenAI's security news is a bit different from the headlines.
The money:
โข $172,000 is the top payout, with bonuses stacked onโจโข A critical bug in ChatGPT, Codex or the API now pays up to $30,000โจโข OpenAI says that's more than four times what critical findings paid before
The partner:
The more interesting part is who they picked. Cathedral is the startup whose AI agents reached Anthropic's internal business systems in October 2026.
Not through Claude.
Through third-party business software.
Anthropic disputes parts of it. It says it took several days, not 12 hours, and that Cathedral went beyond the bounty rules by targeting vendors and not disclosing everything.
Now OpenAI brings them in as a partner, not a bounty hunter. I think that's smart. A bounty comes with a fixed scope and safe harbour. A partner can agree up front what's in bounds.
The Anthropic fight was mostly about scope and disclosure.
What I'd watch:
The path that worked at Anthropic was vendor software, not the model. Bigger bounties for ChatGPT bugs don't cover that.
OpenAI hasn't said what Cathedral will test.
Maybe vendors are already in scope.
If not, the money is pointed at the wrong door.
One line from their write-up on the last sprint stuck with me.
A merged patch didn't always mean the fix had reached every deployed system.
I've seen the same in data pipelines. Code merged, old job still running somewhere.
Links in reply ๐
OpenAI's sprint post: openai.com/index/preparing-oโฆ
Anthropic vs Cathedral dispute: the-decoder.com/ai-agents-haโฆ
Gul Saeed Khattak retweeted
๐ฅ๐๐ป๐ป๐ถ๐ป๐ด ๐ฎ๐ป ๐๐น๐๐ฎ๐๐-๐ข๐ป ๐๐ ๐๐ด๐ฒ๐ป๐ ๐ฌ๐ผ๐๐ฟ๐๐ฒ๐น๐ณ: ๐๐ผ๐ฐ๐ฎ๐น ๐ผ๐ฟ ๐๐น๐ผ๐๐ฑ?
๐ ๐๐ฎ๐ป๐๐ฒ๐ฑ ๐๐ผ ๐ฟ๐๐ป ๐ผ๐ป๐ฒ ๐บ๐๐๐ฒ๐น๐ณ, ๐๐ถ๐๐ต๐ผ๐๐ ๐ฏ๐ถ๐ด ๐บ๐ผ๐ป๐๐ต๐น๐ ๐ฏ๐ถ๐น๐น๐.
The big labs all sell the same idea: ๐ฎ๐ป ๐๐ ๐ฎ๐ด๐ฒ๐ป๐ ๐๐ต๐ฎ๐ ๐๐ผ๐ฟ๐ธ๐ ๐ณ๐ผ๐ฟ ๐๐ผ๐ ๐ฎ๐ฐ/๐ณ, often on its own computer in the cloud.
๐๐น๐ผ๐๐ฒ๐ฑ ๐๐ผ๐๐ฟ๐ฐ๐ฒ
โข OpenAI: Dots
โข Meta: Muse
โข xAI: Grok Bot
โข Anthropic: Claude Duet
๐ข๐ฝ๐ฒ๐ป ๐๐ผ๐๐ฟ๐ฐ๐ฒ
โข OpenMuse (CopilotKit): modelled on Meta's Muse
โข OpenDots (CopilotKit): modelled on OpenAI's Dots
So where do I run it?
๐ ๐ ๐ณ๐ฎ๐๐ ๐ฏ๐ฟ๐ฎ๐ถ๐ป: ๐ญ๐ฐ๐ค๐ข๐ญ๐ญ๐บ, ๐ฐ๐ง ๐ค๐ฐ๐ถ๐ณ๐ด๐ฆ. ๐๐ฉ๐ข๐ต'๐ด ๐ต๐ฉ๐ฆ ๐ธ๐ฉ๐ฐ๐ญ๐ฆ ๐ฑ๐ฐ๐ช๐ฏ๐ต.
๐ ๐ ๐๐น๐ผ๐ ๐ฏ๐ฟ๐ฎ๐ถ๐ป: hang on.
A research agent visits hundreds of websites. Any of them can hide text written for the AI, not for me. That's prompt injection.
It doesn't infect my laptop like a virus. ๐๐ ๐๐ฎ๐น๐ธ๐ ๐บ๐ ๐ฎ๐ด๐ฒ๐ป๐ ๐ถ๐ป๐๐ผ ๐๐๐ถ๐ป๐ด ๐ถ๐๐ ๐ผ๐๐ป ๐๐ผ๐ผ๐น๐ ๐ฎ๐ด๐ฎ๐ถ๐ป๐๐ ๐บ๐ฒ. If the agent has a terminal, one hidden line can say "download and run this."
So I looked at three options.
๐ญ. ๐ ๐ ๐บ๐ฎ๐ถ๐ป ๐น๐ฎ๐ฝ๐๐ผ๐ฝ
Only inside a container or VM, with one shared folder. Free, but the weakest wall.
๐ฎ. ๐ ๐ฐ๐ต๐ฒ๐ฎ๐ฝ ๐๐ฒ๐ฐ๐ผ๐ป๐ฑ ๐บ๐ฎ๐ฐ๐ต๐ถ๐ป๐ฒ ๐ฎ๐ ๐ต๐ผ๐บ๐ฒ
Better. But its security is now my job. And it sits on my home network, next to my router and other devices, unless I put it on a guest network.
๐ฏ. ๐ ๐ฐ๐ต๐ฒ๐ฎ๐ฝ ๐ฐ๐น๐ผ๐๐ฑ ๐ฉ๐ (๐๐ฒ๐๐๐ป๐ฒ๐ฟ, ๐๐๐๐ฟ๐ฒ, ๐ฒ๐๐ฐ.)
Off my home network. Snapshots mean I can wipe it and rebuild in minutes. But I pay for the VM plus the model, and I must not leave its dashboard open to the internet.
Then it clicked. ๐๐ผ๐ฐ๐ฎ๐น ๐๐ ๐ฐ๐น๐ผ๐๐ฑ ๐ถ๐ ๐๐ต๐ฒ ๐๐ฟ๐ผ๐ป๐ด ๐พ๐๐ฒ๐๐๐ถ๐ผ๐ป.
The real question is: ๐๐ต๐ฎ๐ ๐ฐ๐ฎ๐ป ๐๐ต๐ฒ ๐ฎ๐ด๐ฒ๐ป๐ ๐ฟ๐ฒ๐ฎ๐ฐ๐ต ๐๐ต๐ฒ๐ป ๐ถ๐ ๐ด๐ฒ๐๐ ๐๐ฟ๐ถ๐ฐ๐ธ๐ฒ๐ฑ?
A cloud VM holding my Gmail login is as dangerous as my laptop. A laptop agent with one empty folder is fairly safe.
Simon Willison calls the danger the "๐ญ๐ฆ๐ต๐ฉ๐ข๐ญ ๐ต๐ณ๐ช๐ง๐ฆ๐ค๐ต๐ข": private data, untrusted content, and a way to send data out. Remove any one of the three and the data-theft attack stops working.
A research agent needs the untrusted content. So I remove the other two:
โข ๐ก๐ผ ๐ฝ๐ฒ๐ฟ๐๐ผ๐ป๐ฎ๐น ๐น๐ผ๐ด๐ถ๐ป๐ on the agent's machine
โข ๐ฆ๐ฒ๐ฝ๐ฎ๐ฟ๐ฎ๐๐ฒ ๐ฎ๐ฐ๐ฐ๐ผ๐๐ป๐๐ for anything it needs
โข ๐ฆ๐ฝ๐ฒ๐ป๐ฑ๐ถ๐ป๐ด ๐ฐ๐ฎ๐ฝ๐ on every API key (this also fixes the bill fear)
โข ๐๐๐ธ ๐ฏ๐ฒ๐ณ๐ผ๐ฟ๐ฒ ๐ฎ๐ฐ๐๐ถ๐ป๐ด: send, buy, post, delete, run
โข ๐ฆ๐ฝ๐น๐ถ๐ ๐๐ต๐ฒ ๐ท๐ผ๐ฏ: the browsing agent only writes summaries, and the agent with real tools never sees raw web pages
Notice what the big labs did. Muse, Dots and Grok Bot all give the agent a separate cloud computer. Muse keeps credentials where the agent can't read them. Dots stay read-only while idle.
๐ฅ๐๐ป ๐ถ๐ ๐๐ผ๐๐ฟ๐๐ฒ๐น๐ณ, ๐ฎ๐ป๐ฑ ๐๐ต๐ผ๐๐ฒ ๐ฐ๐ต๐ผ๐ถ๐ฐ๐ฒ๐ ๐ฏ๐ฒ๐ฐ๐ผ๐บ๐ฒ ๐๐ผ๐๐ฟ๐.
๐๐ข๐ด๐ต ๐ฃ๐ณ๐ข๐ช๐ฏ ๐ข๐ด๐ฌ๐ด ๐ธ๐ฉ๐ฆ๐ณ๐ฆ ๐ช๐ต ๐ด๐ฉ๐ฐ๐ถ๐ญ๐ฅ ๐ณ๐ถ๐ฏ. ๐๐ญ๐ฐ๐ธ ๐ฃ๐ณ๐ข๐ช๐ฏ ๐ข๐ด๐ฌ๐ด ๐ธ๐ฉ๐ข๐ต ๐ช๐ต ๐ธ๐ช๐ญ๐ญ ๐ฉ๐ฐ๐ญ๐ฅ ๐ธ๐ฉ๐ฆ๐ฏ ๐ช๐ต ๐จ๐ฆ๐ต๐ด ๐ต๐ณ๐ช๐ค๐ฌ๐ฆ๐ฅ.
Links in reply ๐
Gul Saeed Khattak retweeted
๐๐ฒ๐ ๐๐น๐ฎ๐๐ฑ๐ฒ ๐ง๐๐ฟ๐ป ๐๐ป๐ ๐๐ฎ๐ฟ๐ฑ ๐ง๐ผ๐ฝ๐ถ๐ฐ ๐๐ป๐๐ผ ๐ฎ ๐๐น๐ถ๐ฐ๐ธ๐ฎ๐ฏ๐น๐ฒ ๐ ๐ฎ๐ฝ, ๐ช๐ถ๐๐ต ๐๐ต๐ฒ ๐ฃ๐ฟ๐ผ๐บ๐ฝ๐๐ ๐๐๐ถ๐น๐ ๐๐ป
I have seen a better way in Claude recently.
๐๐ต๐ฎ๐ ๐ถ๐ ๐ฎ ๐๐๐ฟ๐ฎ๐ถ๐ด๐ต๐ ๐น๐ถ๐ป๐ฒ. ๐ฌ๐ผ๐ ๐ผ๐ป๐น๐ ๐น๐ฒ๐ฎ๐ฟ๐ป ๐๐ต๐ฎ๐ ๐๐ผ๐ ๐๐ต๐ถ๐ป๐ธ ๐๐ผ ๐ฎ๐๐ธ, ๐ฎ๐ป๐ฑ ๐๐ต๐ฒ๐ป ๐ฎ ๐๐ผ๐ฝ๐ถ๐ฐ ๐ถ๐ ๐ป๐ฒ๐ ๐๐ผ ๐๐ผ๐, ๐๐ผ๐ ๐ฑ๐ผ๐ป'๐ ๐ธ๐ป๐ผ๐ ๐๐ต๐ฎ๐ ๐๐ผ ๐ฎ๐๐ธ.
Ask it for a map of a topic, and it draws a flowchart. The topic is split into steps in a logical order: what comes first, what comes next.
๐๐ป๐ฑ ๐ฒ๐๐ฒ๐ฟ๐ ๐ฏ๐ผ๐
๐ถ๐ ๐ฐ๐น๐ถ๐ฐ๐ธ๐ฎ๐ฏ๐น๐ฒ. Tap one, and Claude writes the next question for you. Usually it sends straight away. Sometimes it lands in your typing box and you just press send. Claude answers that part only, often with a new flowchart. Then you tap again.
๐๐ฒ๐'๐ ๐ง๐ฟ๐ ๐๐: ๐ง๐ฟ๐ฎ๐ป๐๐ณ๐ผ๐ฟ๐บ๐ฒ๐ฟ๐
Say you want to learn how a transformer works. That's the design from the 2017 paper "๐๐ต๐ต๐ฆ๐ฏ๐ต๐ช๐ฐ๐ฏ ๐๐ด ๐๐ญ๐ญ ๐ ๐ฐ๐ถ ๐๐ฆ๐ฆ๐ฅ".
You ask for a map. You get boxes something like this:
โข ๐ง๐ผ๐ธ๐ฒ๐ป๐: your text is cut into small pieces
โข ๐๐บ๐ฏ๐ฒ๐ฑ๐ฑ๐ถ๐ป๐ด๐: each piece becomes a list of numbers, plus its position
โข ๐๐๐๐ฒ๐ป๐๐ถ๐ผ๐ป: each piece looks at the others
โข ๐๐ฒ๐ฒ๐ฑ-๐ณ๐ผ๐ฟ๐๐ฎ๐ฟ๐ฑ: each piece is processed on its own
โข ๐๐ฎ๐๐ฒ๐ฟ๐: attention and feed-forward repeat many times
โข ๐ก๐ฒ๐
๐ ๐๐ผ๐ธ๐ฒ๐ป: the model scores every possible next piece
Attention sounds like the magic part, so you tap it.
You didn't type anything. The question about attention is written for you, and Claude draws a new chart just for that part:
โข ๐ค๐๐ฒ๐ฟ๐, ๐ธ๐ฒ๐, ๐๐ฎ๐น๐๐ฒ: three versions of each piece
โข ๐ฆ๐ฐ๐ผ๐ฟ๐ฒ๐: compare queries with keys
โข ๐ฆ๐ผ๐ณ๐๐บ๐ฎ๐
: turn scores into weights
โข ๐ ๐ถ๐
: a weighted mix of the values
โข ๐ ๐ฎ๐ป๐ ๐ต๐ฒ๐ฎ๐ฑ๐: several attentions running side by side
Tap "Query, key, value", and it goes one level deeper. ๐๐ป๐ฑ ๐ผ๐ป ๐ถ๐ ๐ด๐ผ๐ฒ๐.
๐ช๐ต๐ ๐ ๐๐ถ๐ธ๐ฒ ๐ง๐ต๐ถ๐
โข ๐ฌ๐ผ๐ ๐๐ฒ๐ฒ ๐๐ต๐ฒ ๐บ๐ฎ๐ฝ ๐ฏ๐ฒ๐ณ๐ผ๐ฟ๐ฒ ๐๐ต๐ฒ ๐ฑ๐ฒ๐๐ฎ๐ถ๐น๐.
โข ๐๐ฎ๐ฐ๐ต ๐ฏ๐ผ๐
๐ถ๐ ๐ฎ ๐พ๐๐ฒ๐๐๐ถ๐ผ๐ป ๐๐ผ๐ ๐ฑ๐ถ๐ฑ๐ป'๐ ๐ต๐ฎ๐๐ฒ ๐๐ผ ๐๐ฟ๐ถ๐๐ฒ.
โข ๐๐ณ ๐๐ป๐ด๐น๐ถ๐๐ต ๐ถ๐๐ป'๐ ๐๐ผ๐๐ฟ ๐ณ๐ถ๐ฟ๐๐ ๐น๐ฎ๐ป๐ด๐๐ฎ๐ด๐ฒ: it isn't mine. Phrasing the follow-up takes real effort. A tap skips that.
๐ง๐ต๐ฒ ๐๐ผ๐ป๐ฒ๐๐ ๐ฃ๐ฎ๐ฟ๐
โข ๐๐'๐ ๐ป๐ฒ๐: Claude's in-chat visuals launched in beta in March 2026. Other apps may draw a static image or just text.
โข ๐ง๐ฎ๐ฝ๐ ๐ฎ๐ฟ๐ฒ๐ป'๐ ๐ฐ๐ผ๐ป๐๐ถ๐๐๐ฒ๐ป๐: sometimes the question sends straight away. Sometimes it lands in your typing box and you press send.
โข ๐๐ ๐๐ผ๐ฟ๐ธ๐ ๐ฏ๐ฒ๐๐ ๐ถ๐ป ๐๐น๐ฎ๐๐ฑ๐ฒ'๐ ๐ฐ๐ต๐ฎ๐ ๐ฎ๐ฝ๐ฝ. ๐๐ป ๐๐น๐ฎ๐๐ฑ๐ฒ ๐๐ผ๐ฑ๐ฒ, ๐ถ๐ ๐ถ๐๐ป'๐ ๐ฎ๐ ๐๐บ๐ผ๐ผ๐๐ต.
โข ๐ฌ๐ผ๐๐ฟ ๐ฏ๐ผ๐
๐ฒ๐ ๐๐ถ๐น๐น ๐ฑ๐ถ๐ณ๐ณ๐ฒ๐ฟ: the chart above is an example, not a fixed output.
โข ๐ง๐ต๐ฒ ๐ฟ๐ฎ๐ฏ๐ฏ๐ถ๐ ๐ต๐ผ๐น๐ฒ: four charts later, you can lose where you started. Ask for the path at the top of every chart.
โข ๐ง๐ผ๐ผ ๐ป๐ฒ๐ฎ๐: a flowchart can make a messy topic look simpler than it is.
๐ช๐ต๐ฒ๐ฟ๐ฒ ๐ถ๐ ๐๐ผ๐ฟ๐ธ๐: ๐๐น๐ฎ๐๐ฑ๐ฒ ๐ฐ๐ต๐ฎ๐.
๐ช๐ต๐ฒ๐ฟ๐ฒ ๐ถ๐ ๐ฑ๐ผ๐ฒ๐๐ป'๐: ๐๐น๐ฎ๐๐ฑ๐ฒ ๐๐ผ๐ฑ๐ฒ. ๐ง๐ต๐ฒ๐ฟ๐ฒ, ๐ถ๐ ๐ฑ๐ผ๐ฒ๐๐ป'๐ ๐๐ผ๐ฟ๐ธ ๐ฎ๐ ๐ฒ๐
๐ฝ๐ฒ๐ฐ๐๐ฒ๐ฑ ๐ฎ๐ ๐ฎ๐น๐น.
๐ง๐ฟ๐ ๐๐ ๐ฌ๐ผ๐๐ฟ๐๐ฒ๐น๐ณ (๐ฆ๐๐ฎ๐ฝ ๐๐ต๐ฒ ๐๐ฟ๐ฎ๐ฐ๐ธ๐ฒ๐๐)
Give me a clickable map of [topic] as a flowchart. Use 5 to 7 boxes, in the order I should learn them. Make every box send its follow-up message as soon as I tap it, so I never type anything or press send. That message should ask about that step only and draw a new flowchart for it the same way. Show the path at the top of every chart (like: Transformer > Attention > Query, key, value) so I don't get lost.
๐๐ต๐ฐ๐ฑ ๐จ๐ถ๐ฆ๐ด๐ด๐ช๐ฏ๐จ ๐ต๐ฉ๐ฆ ๐ฏ๐ฆ๐น๐ต ๐ฒ๐ถ๐ฆ๐ด๐ต๐ช๐ฐ๐ฏ. ๐๐ด๐ฌ ๐ง๐ฐ๐ณ ๐ต๐ฉ๐ฆ ๐ฎ๐ข๐ฑ.
Links in reply ๐