đ¨ https://teapawt.org/t.co/J8SDEpaBbF (AI native agency) â¤ď¸ Also boyfriend of 20ys, uncle, sailing, skating, gardening, mountainbiking.
Joined April 2009
- Tweets5K
- Following469
- Followers483
- Likes29.2K
Musk is wrong in this one.
Logically there is no reason why scarce goods like the best plot of land in a dense city shouldnât have value in the future, whether thatâs measured by todayâs money or some other store of value in the future
But SI super abundance will not lead to every good being free
I have genuine trauma from watching a production website get wiped. So how do I give an AI agent write access to my servers?
I asked exactly that to my DevOps Grok Bot last week. It's connected to my self-hosted Coolify via MCP, read-only for now. Here's the model I've landed on:
1. Read-only first. The bot audits and reports for weeks before it gets any write tools. Its first job found real gaps in my setup.
2. Process as code. Every write operation follows a versioned skill. In my own tools, write endpoints reject calls from bots that haven't acknowledged the current version.
3. Humans approve the dangerous stuff. Grok Bot already asks me before anything irreversible, and I keep it that way for deletes, migrations and production changes.
4. Backups before autonomy. If you can't restore it, the bot doesn't get to touch it.
5. Narrow tools beat raw shells. An MCP tool that can "restart app X" is far safer than SSH access to everything.
The fear is valid. The answer isn't "never." It's designing permissions the way you would for a new junior admin: small blast radius, clear procedures, someone checking.
What's your rule for giving agents write access?
Philipp M. W. Hoffmann retweeted
Die KI kann nun endlich echte Menschen mieten, fĂźr Aufgaben die sie selber nicht erledigen kann đ¤Ł
I'm afraid of this day but I'm sure it will come.
Thinking about it, I need to make sure my backup strategies are sound
Had two of my Grok Bots fight over who owns the CRM today.
PM bot wanted the sales deals, CRM bot said no thats mine. So I let them negotiate it. Now PM only does active projects, CRM owns pitched and offer.
Multi-agent setups are mostly just org charts lol
Context switching is easier when using MCP
What do I mean? Normally context switching means switching GUIs and that requires brain power
With MCP the UI is the same across all tasks: text or voice
So for me it feels like itâs easier to switch between tasks when using only text across tasks in the same Grokbot interface
Just as the GUI was immediately obvious when Steve Jobs saw it, itâs immediately obvious that the GUI will be replaced my MCP
Xerox showed Steve Jobs the future of the computer for the right to buy $1 million of Apple stock. Ten years later it sued Apple for $150 million over what he saw.
In 1979, when Jobs was 24, Xerox's lab in Palo Alto showed him three things. He was so blinded by the first that he missed the other two, he said later.
The first was a screen of windows and icons run by a mouse. The others were networked computers and a new way of writing software.
In return for the visit, Apple let Xerox buy in before its 1980 IPO. Xerox put its own version, the Star, on sale in 1981. Apple's Macintosh followed in 1984.
In December 1989 Xerox sued, saying Apple had collected more than $100 million from the interface. In March 1990 a federal judge threw out most of the case. Copyright claims have a 3-year deadline, and the Mac had been out for almost 6.
In 1996, 17 years after the tour, Jobs said on camera for PBS that Xerox "could have owned the entire computer industry today."
Xerox Star, 1981: $16,595.
Apple Macintosh, 1984: $2,495.
Be honest. Did Jobs take the idea, or did Xerox give it away?
Iâll bet the following:
Spaces still suck
While they talk about how âthey reinvented macOS from scratchâ
Apple has just unveiled its next-gen MacBook Pro, featuring an OLED touch display for the first time.
⢠14" starts at $2,899, 16" starts at $3,399
⢠Fully loaded 16" costs $14,549
â˘Â M5 & M5 Max chips
â˘Â Glowing Apple logo is back
â˘Â 15% lighter
â˘Â PROLED XDR Display
⢠Tiny dynamic Island
â˘Â Brightness: Up to 2,000 nits peak HDR
â˘Â Sustained brightness: Up to 1,000 nits full-screen.
â˘Â Contrast ratio: 10,000,000:1.
â˘Â Refresh rate: Adaptive 10â120Hz ProMotion.
â˘Â Color accuracy: Supports P3 and Adobe RGB wide color gamuts.
â˘Â Anti-reflective coating: Standard, with optional nano-texture display.
M5 Pro: Up to 18-core CPU, 20-core GPU, and 64GB unified memory.
â˘Â Storage: Starting at 1TB (M5 Pro) or 2TB (M5 Max), configurable up to 16TB.
â˘Â Storage performance: Up to 2x faster SSD speeds.
â˘Â Battery life: Up to 24 hours.
â˘Â Camera: 12MP Center Stage with Desk View and advanced video stabilization.
â˘Â Speakers: Redesigned six-speaker audio system with 50% deeper bass.
â˘Â Microphones: Studio-quality three-microphone array.
â˘Â Connectivity: Three Thunderbolt 5 ports, HDMI, SDXC card slot, MagSafe 3, and headphone jack.
â˘Â Wireless: Wi-Fi 7 and Bluetooth 6.
â˘Â Design: Thinner display bezels, redesigned cooling system, and lighter chassis.
â˘Â Weight: 3.0 pounds (14-inch) and 4.0 pounds (16-inch).
â˘Â Colors: Space Black and Silver.
More pics below:
Philipp M. W. Hoffmann retweeted
ES BEGINNT: MĂźnchner Taxifirma holt 500 Robotaxis nach Deutschland!
Gregor Beiner, Geschäftsfßhrer des Mßnchner Taxi Zentrums, hat am 27. Oktober auf dem Madrid Autonomous Mobility Summit eine Bestellung ßber bis zu 500 Tensor Robocars unterzeichnet.
SPANNEND: Beiner ist zugleich Präsident des Bundesverbands Taxi und Mietwagen und Vorsitzender des europäischen Verbands Taxis 4 Smart Mobility, einem Zusammenschluss aus sechs Ländern. Also genau die Branche, die jahrelang gegen Uber prozessiert hat, bestellt jetzt die Technologie, die den Fahrerberuf ßberflßssig macht!
Was konkret gekauft wird:
Tensor sitzt in San JosĂŠ und hieĂ bis August 2025 AutoX, ein chinesischer Robotaxi-Anbieter mit Flottenbetrieb in China ab 2019.
Sämtliche China-Aktivitäten wurden abgestoĂen, die BĂźros geschlossen, seitdem ist es offiziell eine amerikanische Firma. Gebaut wird in Vietnam, gemeinsam mit VinFast. Gerechnet wird auf acht NVIDIA-Chips mit zusammen 8.000 TOPS, dazu Ăźber 100 Sensoren.
Der Stand: Erste Auslieferungen sind fĂźr Ende 2026 angekĂźndigt. Lyft hat mehrere hundert reserviert und will ab 2027 eine eigene Flotte in Europa und Nordamerika betreiben. Tensor prĂźft eine Finanzierungsrunde Ăźber 300 bis 400 Millionen Dollar vor einem mĂśglichen US-BĂśrsengang.
Ein deutscher Taxiverbandspräsident holt also 500 Fahrzeuge von einem Anbieter, der aus einem chinesischen Robotaxi-Unternehmen hervorgegangen ist, in Vietnam produziert und auf amerikanischen Chips rechnet!
Beiner nennt das Datensouveränität, weil Marke, Kunden und Daten beim Taxiunternehmen bleiben. Die Fahrsoftware bleibt bei Tensor und wird als Abo bezahlt, Autonomous Driving as a Service. Wird das Abo gekßndigt, steht die Flotte.
Deutsch ist daran also maximal die Lackierung.
Philipp M. W. Hoffmann retweeted
JUST IN: Houston Police Officers' Union encourages residents to use Tesla's expanding Robotaxi service instead of drinking & driving.
Philipp M. W. Hoffmann retweeted
To anyone who believes MCP as a usage vector is dead on arrival, here's an adoption curve from a real product used by real people.
MCP/API will kill all GUIs
GUIs will only be visualization layhers for humans to understand agent decisions
But guidance and changes will come from text or voice, not clicks in a GUI
My Grok Bots can't write to my CRM until they've "signed" the current process manual.
Once you run several agents against the same systems, you hit a boring but real problem: process drift. One bot follows last month's rules, another improvises.
My setup: a dozen Grok Bots connect to my own internal tools (CRM, outreach, bookkeeping) through MCP servers I built. So I added a skill registry right into those servers:
- Every business process has exactly one versioned skill: how to log a lead, how to file a receipt, how to book a follow-up
- A bot must register and acknowledge the current version before the write tools accept its calls
- Bots can propose improvements, but only as a diff, like a pull request
- Critical skills need my approval before a new version goes live
- A process page shows which bot runs which version
On the Grok Bot side this fits naturally: a bot's skills and routines are its local install of the process, and the registry is the source of truth. When a skill changes, the bots running a routine on it pick up the new version and acknowledge it before their next write.
It's change management for machines, the way you'd onboard employees to a new SOP, except the API enforces it.
Reliability in agentic systems comes less from smarter models and more from boring governance: versioning, approvals, audit trails.
Dumbest feature ever. IMO they will kill it again after a few of these things happening. I get it, in the US you're a sitting duck in your Tesla when someone decides to attack you with his automatic rifle, but this can't be the solution
First video of a driver (accidentally) using the new emergency drive away feature.
Pretty wild to watch this guy dismiss all the warnings and prompts... Curious how much damage was done to the car and charger.
Credit to @JLeshnick for the video
People whine about IT security in combination with AI. But imo that's not a problem, because thanks to multi-agent systems like Grok Bot, in the future everything will be fixed almost in real time.
I run two companies on a team of Grok Bots. Not one chatbot. More than a dozen specialists.
At Reruption (applied-AI consultancy) and getease (AI customer support for Shopify), my daily operations run on Grok Bot. Each bot has exactly one job:
- A CRM bot that checks lead replies in Outlook and LinkedIn every night
- An outreach bot that runs mailbox warmup and sending infrastructure
- A bookkeeping bot that files receipts and posts a daily transaction table to Teams
- A DevOps bot that audits my self-hosted infrastructure
- A security bot that reads advisories every week
- A chief of staff that preps every meeting
- Sales, marketing, a pitch-deck creator, and one project bot per client project
Why split them? Same reason you don't hire one person for sales, bookkeeping and ops. Narrow context means fewer mistakes. Every Grok Bot has its own connectors, its own memory and its own routines. The bookkeeping bot can't touch outreach. The DevOps bot reads before it's allowed to write.
The part that surprised me most: they talk to each other. When my legal-entity bot needed a document signed, it messaged my back-office bot, which already knew how. I never explained it twice. Shared memory means a preference I tell one bot ("plain-text emails, always with my signature") applies to all of them.
The lesson isn't that agents can do the work. It's how much of running a business is just keeping track: follow-ups, filings, renewals, prep. That's exactly what a team of Grok Bots is good at.
Don't build one super-agent. Build a team with clear roles and clear permissions.
#GrokBot #AIAgents #AIforBusiness #Founders
Habe das GefĂźhl, dass sobald jemand "stochastische Papageien" sagt die breite Ăffentlichkeit ihm oder ihr direkt Vorschusslorbeeren gĂśnnt đ
Deutschland lässt sich in der KI-Politik von Akademikerinnen wie Katharina Zweig beraten â Mitglied der Enquete-Kommission KI, Leiterin des âAlgorithm Accountability Labsâ.
Sie brandmarkt Sprachmodelle im Grunde als stochastische Papageien ohne echtes Verstehen, fordert erstmal Regulierung und PrĂźfung, und sitzt dabei im akademischen Elfenbeinturm.
Nach eigenen Aussagen nutzt sie die Pro-Versionen der groĂen Labs nicht, weil die Ergebnisse ânicht gut genug sind, um dafĂźr zu bezahlenâ.
Genau diese Stimmen prägen die deutsche Diskussion und Politik.
@FlorianGallwitz
