I make things: code, songs, AI art & books Data Scientist | Full Stack + Web3 Singer • Songwriter • Amazon Author Boxing through my 20s | Learning in public
📍Earth
Joined October 2016
- Tweets59K
- Following951
- Followers5.2K
- Likes69.7K
Pinned Tweet
Absolutely speechless.
For the second time, @elonmusk has liked one of my replies.
I get my share of unnecessary abuse as a small creator, but I've learned to just ignore the noise.
As a small account, you just post your thoughts, never expecting them to reach the person you're talking about. This is a dream.
It's moments like this that make it all worth it. Made my entire week.
Thank you Elon.
I am emotional , i cant express my feeling.
Myth: a SpaceX and Tesla merger is basically done.
Reality: Musk said last night that would be overstating it.
What he confirmed on SpaceX's earnings call:
• Both companies formed special committees of independent directors at shareholders' request
• He can't take part because he's a conflicted party, and called it an odd kabuki dance
• The committees have not agreed on any terms, and there's no guarantee they will
The biggest hurdle is the vote.
On the Tesla side, a deal would need 50% of all shares, not just of shares that vote. When turnout is often well below 100%, that's like needing a majority of every registered voter, not just those who show up.
Musk still called it a good idea. He pointed to shared work on AI software, chip design, Terafab, robotics and the Starmind constellation.
About Dec 3: that's Tesla's annual meeting, at 2:30 AM IST on Dec 4. The proxy covers director elections, a pay vote and the auditor. There's no merger vote on it.
Barclays suggested a deal could be announced before that meeting. Any merger would still need its own shareholder vote.
A combined company could be worth over $3.6 trillion, per The Verge.
The idea has momentum.
The paperwork doesn't exist yet.
Why would a fab built for AI chips start with memory?
Terafab's first chip, due around May or June 2027, is an SRAM test chip. Musk said it basically tests the line.
The problem: a modern chip passes through hundreds of process steps. If anything goes wrong, a full processor won't tell you where.
Here's how an SRAM test chip works:
▷ SRAM is the fast memory inside every processor, built from a tiny repeating cell of six transistors
▷ A test chip packs millions or billions of identical cells, the densest and most demanding pattern a process can print
▷ Engineers write data in, read it back and map every cell that fails
▷ The failure map is a diagnosis. A cluster in one corner points to a tool or an uneven layer. A scattered pattern points to particles or a chemistry problem
Think of it like a thermostat. You can't tune the heating if you can't read the temperature.
Yield on SRAM becomes the scorecard. Fabs push it up week by week until it's high enough to risk complex logic.
Reuters noted that chipmakers commonly use SRAM circuits this way to refine a new process.
That's why the test chip comes roughly a year before Terafab's volume target of Q3 2028, which Musk himself called very aggressive.
A test chip doesn't sell.
It tells you whether anything else will.
Starlink wants to be your carrier. Low band spectrum is how.
On last night's call, Shotwell said the new spectrum will pave the way for Starlink Mobile to become a major mobile carrier in the US.
The pieces:
▷ On Oct 8, SpaceX agreed to buy Grain Management's nationwide 800 MHz portfolio, up to 14 MHz of paired spectrum. It still needs FCC approval
▷ The FCC approved a constellation of 15,000 Starlink Mobile V2 satellites a few weeks ago
▷ Those satellites start launching on Starship in 2027
Why 800 MHz matters: low frequencies travel farther and pass through walls and trees better. It's why your signal often survives inside a building on low band and dies on high band.
Shotwell expects S band spectrum to carry most of the traffic, with low band carrying less but filling a critical gap.
She also pointed out that somewhere between 85% and 93% of mobile data already goes over WiFi.
The scale today:
• 22 million monthly Starlink Mobile users, up 37% in a quarter
• New carrier partners in Uganda, the Democratic Republic of Congo, Kazakhstan and Costa Rica
Separately, the 800 MHz licenses still need their own FCC approval for satellite use.
Towers cover where people live.
Satellites cover everywhere else.
The metal that saved Moore's law is named for Copenhagen.
Hafnium, from Hafnia, the Latin name for the city, sits in almost every advanced chip you own.
How it got there:
→ For about 40 years, chips used silicon dioxide as the thin insulating layer in every transistor gate
→ By the mid 2000s that layer was only a few atoms thick, and electricity started leaking straight through it
→ Leakage meant heat and wasted power, and shrinking further looked impossible
→ In 2007, Intel shipped its 45nm Penryn chips with hafnium based high k dielectric and metal gates
→ Gordon Moore called it the biggest change in transistor technology since the late 1960s
Why hafnium oxide works: high k means it stores charge far better than silicon dioxide. So engineers can make the layer physically thicker, which blocks leakage, while it still acts electrically thin.
Picture a dam that holds back more water by being wider, not taller.
Every major chipmaker followed. Today hafnium oxide is standard in leading edge logic, and it's the kind of material a new fab like Terafab has to master on day one.
It even has a second life. A form of hafnium oxide turns out to be ferroelectric, which researchers are using for new kinds of memory.
One metal fixed the leak.
Two decades of shrinking chips followed.
How long can Tesla spend this fast?
Tesla reports Q3 results today, with the call at 3:00 AM IST tomorrow, Oct 22.
The number to watch is cash, not cars. Tesla's own compiled analyst consensus expects:
• Free cash flow of about minus $4.51B for the quarter. A year ago it was nearly plus $4B
• Capex of about $7.77B in the quarter, close to $85M a day
• Revenue of about $28.3B
For the full year, Tesla has guided to more than $25B in capex, versus roughly $8.5B last year. That's about three times as much.
Where it's going: Optimus production lines, AI compute, the Cybercab ramp and the Terafab chip plant with SpaceX.
Then there's the regulator.
NHTSA sent Tesla a special order with 21 questions on how it self certified the Cybercab, which has no steering wheel or pedals, as meeting federal safety standards. Tesla asked for an extension to Nov 20. NHTSA gave 30 days instead, so answers are due Oct 30, nine days after this call.
Cybercab started paid rides in Austin on Sep 3, and Tesla said early this month it had more than 300 in service there.
Musk posted over the weekend that Optimus and Cybercab will be the vast majority of Tesla's value.
The quarter shows the bill.
The call has to show what it buys.
Claude Duet vs every competitor
Claude Duet just joined the always on agent race.
Anthropic opened the research preview and it lands in a crowded field.
Here's how the five stack up:
→ Claude Duet by Anthropic runs on Opus 5.5. Waitlist only, included in Pro and Max, US only. iOS, Android and claude.ai. It sends a morning brief, checks your apps every few hours and works on its own cloud computer with a browser. Over 6,000 connectors, your Claude memory and voice messages.
→ dots by OpenAI run on GPT‑6 Astra. Launched Sept 29. The first dot is included in Pro and Business Premium at no extra cost, with an Enterprise beta. Pro excludes the EEA, Switzerland and the UK. Its own cloud computer, over 4,000 apps, ChatGPT memory, voice calls, plus Slack and Teams.
→ Grok Bot by SpaceXSI launched in beta Aug 11. It's included on The X Subscription Premium at $30 a month, Super at $100 and Ultra at $200. A shared cloud computer with browser and terminal, many Bots working in parallel, and routines it learns by watching you once.
→ Muse by Meta runs on Muse Spark. Launched Sept 8 in the US. Free with a weekly limit, or $20 and $100 plans. iOS, Android, web and WhatsApp. Its own isolated VM, side chats, a Goals tab and memory files you can edit.
→ Gemini agent by Google launched Oct 8 in private preview for select enterprise customers. It picks between Gemini and Claude models per job.
The guardrails differ:
• Duet asks before it sends, pays, signs or deletes, and refuses securities trades and transfers between accounts even with approval
• dots research in read only mode and follow your custom rules
• Muse routes every action through a Sentinel the agent can't override and pays with single use cards
• Duet, dots and Muse all keep passwords away from the model
Duet arrived last and claims the biggest connector count of anyone who published one.
Now it has to prove its judgment.
4. Muse and Gemini
Meta Muse gives you the most control over the assistant itself:
• A dial to turn its proactive messages up or down
• Side chats and a Goals tab
• Memory saved as files you can read and edit
• Weekly token caps on every plan
Google's Gemini agent works across Workspace, Microsoft 365, Slack and a CLI, with governance and cost controls for companies. Google plans to include it with Gemini Enterprise.
Five agents. Five ideas of what an assistant should be.
3. Grok Bot's different bet
Grok Bot isn't one assistant. It's a team you build.
▷ It launched on desktop and iOS
▷ Usage comes from one weekly pool shared with Grok and Cursor, with top ups after
▷ It gets more proactive the longer you use it
One bot for everything versus many bots for each job.
That's the real split in this race.
2 / Where dots goes further
OpenAI's dots reach more places than any rival today:
• ChatGPT web, desktop including Windows, iOS and Android
• Texting is in beta, so a dot can reach you without an app
• It can use your local computer as well as its own cloud one
• A safety monitor can pause it mid task
The rollout is gradual, so not every Pro user has it yet.
1 / Duet's fine print
The details that didn't fit in the main post:
→ Duet can't show you its browser while it works, so you see results, not the screen.
→ Background checks count against your Pro or Max usage limits.
→ When it does pay with your approval, it uses a Link virtual card rather than your real card
Anthropic says it can still misjudge what's urgent and will keep tuning it from early user feedback.
I didn't see this coming: Anthropic just joined the always-on agent race with Claude Duet.
Claude Duet is a proactive AI assistant powered by Opus 5.5 that works in the background, even when your phone and laptop are off.
Here's what it does:
→ Sends you a short brief every morning on what matters across your apps
→ Checks your apps every few hours and only pings you when something needs you
→ Takes tasks by text or voice in one ongoing thread, from booking meetings to building dashboards
→ Runs on its own cloud computer with a browser
It carries over your Claude memory, skills and connected apps, with access to more than 6,000 connectors including Gmail, Slack, Stripe and Shopify.
It still asks before it sends, pays, signs or deletes anything, and it never directly sees your passwords.
Right now it's a research preview on a waitlist for Pro and Max users in the US.
The chatbot era was about asking AI for help.
The agent era is about AI showing up before you ask.
Claude Duet is an early research preview, available through a waitlist on Pro and Max plans in the US. It can still misjudge what’s urgent, and we’ll keep tuning it based on what early users tell us.
Join the waitlist: claude.com/product/claude-du…
Read more: claude.com/resources/article…
CLAUDE NOW COMES TO YOU FIRST.
At 12:39 AM IST today, Anthropic opened a research preview of Claude Duet, a new app where Claude works as a proactive agent.
You don't open a chat and wait. Duet messages you.
How it works:
→ Each morning it sends a short brief on what matters across your connected apps
→ It checks those apps every few hours and pings you only when something needs you
→ You hand it tasks by text or voice in one ongoing thread, from booking a meeting to building a dashboard or forecast
→ It runs on its own cloud computer with a browser, so it keeps working with your phone and laptop off
It's built around Claude Opus 5.5 and carries over your memory, skills and connected apps from Claude. It can reach more than 6,000 connectors and plugins, from Gmail and Slack to QuickBooks, Stripe, Gusto and Shopify.
The guardrails:
• It asks before it sends a message, pays, signs or deletes anything
• Trading securities and moving money between accounts are off limits even with your approval
• It never directly sees your passwords, and a new classifier blocks sharing card numbers and government IDs
Who gets it: Pro and Max subscribers in the US, through a waitlist. It's included in those plans and shares their usage limits, and background checks count too. Team and Enterprise aren't included yet. Apps on iOS, Android and claude.ai.
Anthropic says it can still misjudge what's urgent.
Chatbots wait for you.
Duet is betting the best assistant doesn't.
How to get it
Claude Duet is an early research preview.
It's open through a waitlist for Pro and Max subscribers in the US, with no extra cost beyond the plan. Team and Enterprise aren't included yet.
It works on iOS, Android and claude.ai.
Anthropic admits it can still misjudge what's urgent, and it will keep tuning Duet based on what early users say.
Join the waitlist at claude.com/product/claude-du…
Always on
Duet doesn't live on your phone.
It runs on its own cloud computer with a browser, so it keeps working while your laptop and phone are off.
▷ A short brief every morning
▷ A check of your apps every few hours
▷ A ping only when something truly needs you
It connects to more than 6,000 apps and plugins, from Gmail and Slack to Stripe, QuickBooks and Shopify.
Duet builds on years of Anthropic's safety and security work on agents.
Here's what protects you:
• It never directly sees your passwords
• A new classifier blocks it from sharing card numbers and government IDs
• It asks before it sends, pays, signs or deletes anything
• Trading securities and moving money stay off limits, even with your approval
Anthropic says it will share more on its safety approach soon.