Sam Altman@sama
6.1 Terra is here! It is good at a lot of things, but we are especially excited for this new era of computer use.
OpenAI@OpenAI
GPT-6.1 Terra: delivering fast and capable computer use at a lower price.
AI Showcase
Jinhao Hu
How Jinhao works with AI
I define the work. AI crews do it. Work I flag for acceptance counts only after an independent review.
01 · The crew
Parallel crew sessions, each on its own task.
Each pane is one crew session in its own terminal. Cyan is working, amber is waiting for review, grey is idle. Tap or hover a pane for the details.
Example task
- Task
- Current step
- Output
- Review status
02 · AgentDeck
One captain, many hands, one reviewer.
A desktop app where I talk to one lead agent, the captain, which hands work to crew sessions running in parallel.
- 01IntentOne captain. I talk to it; it dispatches.
- 02DispatchThe captain splits the work across parallel crew sessions.
- 03Parallel workEach crew session works in its own terminal and logs what it does.
- 04Results returnResults come back with a receipt.
- 05ReviewThe independent reviewer checks files, commits and tests. Failed work goes back automatically.
- 06BoardWork I flag for acceptance moves to done only after review. I set the rule and approved the checks.
- 07Back to meWork that passes review comes back to me. I accept it or send it back.
MeIntentAccepted ✓“Summarize this week’s lectures and flag what’s due.”
Lead agentCaptain
coursesread 3 slide decksfound 2 deadlines
intelgrouped 41 posts5 themes
ai-newsfetched 12 headlinesdrafting summary
stock-researchsummarized 3 filingsresearch only
reading-listtagged 9 articlessorted queue
side-projectpatched layoutrunning tests
IndependentReviewerReview required
Task boardDone0
Sent back: missing test
Back to me: I accept it or send it back
Back to me ↑Sign-offWork that passes review comes back to me. I accept it or send it back.
Illustrative animation · the numbers below are real
- 1,471agent tasks completed, across all my projectsTask board · cards created Oct 4 to 24, done by Oct 28, 2026
- 145 of 1,505agent tasks sent back for reworkTask board · Oct 4 to 24, 2026
- 5task states on the boardAgentDeck task board design
- 124automatic captain handoffs in 21 daysHandoff log · Oct 4 to 24, 2026
02 · One real task
A real card, sent back once.
Priority marker: show at a glance which tasks are urgent. Step through it: the request, the card, the delivery, the review that sent it back, the rework and the pass.
Real task card from my board. Translated from Chinese and shortened; file paths removed.
MeRequest
By “high priority” I mean work the captain should start right away in the background. I want a clear mark on the task board or the sidebar, so I can see at a glance which task is urgent.
Captain (lead agent)Task card
The captain can mark a card’s priority by command, when creating it and later. The board card, the sidebar session and the phone web page all show it; high priority sorts first. UI work: wait until the chat-page spacing task is done, so two sessions never edit the same UI files.
Crew sessionDelivered
Branch pushed (commit cbc958f) with a report and 14 screenshots.
Independent reviewerSent back
Not passed. The high-priority flag stays on the crew session after the task completes. I loaded the real session code in an isolated runtime: give it an urgent task, complete it, then send an ordinary, unrelated task to the same crew session. The new task comes out as high priority, so it would jump the sidebar order. Fix the priority lifecycle and add a regression test for exactly this. Checked: report and 14 screenshots present, commit on the remote, related unit tests 197/197 and end-to-end tests 33/33 pass, no test deleted or loosened. I changed no code.
Crew sessionRework
Rework pushed (commit 9ce020a). The high-priority mark now lives on the task’s dispatch record instead of the crew session, so an ordinary task sent after completion no longer carries the flag. Added 4 regression tests that fail before the fix and pass after. Unit tests: 1,217 run, 1,204 passed, 13 skipped, 0 failed. End-to-end: 81 run, 79 passed, 1 failed (it fails on the untouched baseline too), 1 not run. Screenshots redone.
Independent reviewerPassed
Passed. Every file and all 14 screenshots checked; commit 9ce020a exists and matches the remote; the diff deletes or loosens no test assertion; I ran the 19 tests for this change and 153 related core unit tests myself, all passed.
Task boardDone
This card was flagged for acceptance, so it moved to done only after the reviewer passed it. It closes with the reviewer’s receipt.
03 · Hermes
A day of scheduled AI jobs.
One private site that brings my email, AI news, videos, podcasts, reading and calendar together on my iPhone, iPad, Mac and Windows PC. Scroll and the clock runs from midnight to midnight; each job lights up at its real time and the phone shows what it sends. Feeds come from scheduled jobs of the open-source Hermes agent, which I use and integrate.
00:00PT
Real schedule · Illustrative output00:00NightNothing is scheduled yet. The first job runs at 01:00.
- 86scheduled jobs enabled out of 118 configuredHermes job list · Oct 28, 2026
- 41of them call an AI agent; 45 are plain scriptsHermes job list · Oct 28, 2026
- 434 of 447AI-job runs succeeded over 13 days (97%)Hermes run log · Oct 12 to 25, 2026
- 3×a day, AI turns new email into action itemsHermes schedule · 9:00, 12:00, 22:00
04 · Fuqing, the real world
A real store, a real system.
A family building-materials store with 50 staff ran on handwritten sales slips, and stock was barely recorded. I built a web inventory system with AI coding agents: stock ledger, credit-sale balances, customer search by phone, and an assistant that turns daily Excel sheets into orders.
Screens in preparation
Screens in preparation
Screens in preparation
Real interface. The screenshots use demo data; the numbers below come from the live store system.
- 2,167orders logged since August 2026Store database · Oct 24, 2026
- 97%entered through the AI Excel import (1,859 of 1,914)Store backup · Oct 16, 2026
- 50staff at the storeThe store, 2026
05 · XDeck
Your X feeds, side by side.
A free desktop app for Windows and Mac that puts five or six X (Twitter) feed columns in one window, without X’s paid multi-column subscription or the paid X API. I set the requirements and Claude wrote the code; the first simple version took about 30 minutes. It is open source on GitHub.
XDeck DIYSam Altman@samaSam Altman@sama
I am very grateful Aleksander accepted our request to come take this on; the safety and security challenges in front of us as we confront superintelligence are immense, and he has an excellent set of skills to work on this.
7883117,1191.46MSam Altman@sama
The back of the Navier-Stokes milestone shirt is the last line of the chain-of-thought from the solution:
Elon Musk@elonmuskElon Musk@elonmusk
Cool
Sundar Pichai@sundarpichai
A ‘hello world’ milestone for Project Suncatcher 🚀 After launching our prototype satellite a few weeks ago, we’ve now successfully sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth - a haiku!
Ground: Write a haiku from the perspective of a TPU orbiting Earth, with the wonderful world spinning beneath you.
Gemma:Blue marble below,Data streams in silent grace,Starlight feeds my core. ✨
Anthropic@AnthropicAI
anthropic.comEarly progress on accelerating drug development with Claude
Google@Google
blog.googleGemini 4 Carbon: our next frontier modelOpenAI@OpenAI
openai.comHow we detect frontier model misalignmentAnthropic@AnthropicAI
anthropic.comDonating Agent Skills to the Agentic AI Foundation
Mistral AI@MistralAI
mistral.aiIntroducing Mistral Large 4
Demis Hassabis@demishassabisPinned
Demis Hassabis@demishassabis
No.1 💎
- 1 · Trending in TechnologyGemini 4 Carbon
- 2 · Trending in TechnologyModel Spec
- 3 · Trending in TechnologyAgent Skills
- 4 · Trending in TechnologyMistral Large 4
- 5 · Trending in TechnologyAI for Science
- 6 · Trending in TechnologyFrontier model misalignment
Real posts captured Oct 30 · samples only in Search and NotificationsTry it: slide the columns sideways, move one with its arrows, drag an edge, or pick how many fit.
- One sign-inLog in once; every column shares the session.
- ReorderShift a column left or right without reloading it.
- ResizeDrag a column edge from 240 to 900 px; widths are remembered.
- Back and forwardEach column keeps its own history, mouse side buttons too.
In the app each column is the real x.com page, signed in once. Here, the Sam Altman, Elon Musk and Demis Hassabis columns are real posts I captured from my XDeck on Oct 30, 2026: their words as posted, counts as of that day, profile pictures theirs. The AI labs column links to the labs’ real articles. Search and Notifications are made-up sample accounts.
- 14GitHub starsGitHub · Oct 25, 2026
- 82release-file downloads across 4 releasesGitHub releases · Oct 25, 2026
- ~30 minto the first simple version (my estimate)My estimate
06 · Résumé AI
A résumé that can’t make things up.
A one-page résumé for a job posting, written only from a fact bank I confirmed. Then another company’s model reviews it blind. I defined the pipeline and its review rules; AI coding agents built it. Contact details are stripped before any model call, and every reviewer finding became a general rule, not a one-off fix.
The needs are read from the job posting.
Projects
- Built a web inventory system with AI coding agents for a 50-staff store: 2,167 orders logged since August 2026.F1F10
- Defined AgentDeck, a desktop workspace that runs many AI coding agents side by side: a peak of 23 sessions at once.A1
- Led a team of 12 engineers to launch an AI platform.No fact in the bank: removed
- Set up 86 scheduled jobs on the open-source Hermes agent; 97% of AI-job runs succeeded over 13 days.H2H8
- Turned 1,664 emails into action items, Sep 16 to Oct 25, 2026, under a prompt that forbids sending mail.H7
Illustrative replay with a sample posting. The page lines are real facts from this site; the reviewer verdicts are examples; the numbers are real.
- 15 of 20first drafts passed both blind reviews on job postings it had never seenTest run · Oct 25, 2026
- 10real job postings across 5 role types used to set the rulesRule set · Oct 2026
- 3 → 4contribution score when my own part comes before the team result (A/B on two pages)A/B test · Oct 2026
07 · Incidents
When the AI got it wrong.
Three real incidents from my own projects: what went wrong, what I did, and what changed.
A half-typed message went out
- What went wrong
- An auto-delivery bug typed a crew session’s receipt into the captain’s input box and pressed Enter, sending my half-typed message with it.
- What I did
- Set a human-in-the-loop rule: typed input is sent only when a person presses Enter.
- Result
- Crew sessions fixed the delivery bug in commit e6d96d75.
Empty work, reported as finished
- What went wrong
- The model setting was dropped when tasks were handed out, so 6 crew sessions ran on Claude Haiku and reported finished work that was empty.
- What I did
- Required every crew session to confirm which model actually ran (Oct 9).
- Result
- The model check has been applied 16 times since.
Discounts booked as payments
- What went wrong
- The old AI import step had booked discounts as payments on orders it brought in from the store’s Excel sheets.
- What I did
- Proposed a row-by-row check of the AI-imported orders against the source sheets. The AI ran the comparison, then scanned the whole database.
- Result
- The scan listed 118 orders with the same problem. I approved the fix and all 118 were corrected on Oct 17.
DefineDispatchReviewShip
AI writes the code. I own the decisions.
08 · How I work
Define, dispatch, review, ship.
- 01DefineI write what the system should do and what done means.
- 02DispatchThe captain routes each task to the model that fits it.
- 03ReviewThe independent reviewer checks it. Two failed rounds suspend the task.
- 04ShipWork I flag for acceptance is merged and installed only after it passes review.
- 175rules in the playbook the captain and every crew session read firstRule book · Oct 24, 2026
- 19 of 19emails sent by the captain and crew sessions in a month went only to me, after my approvalSent mail · Sep 24 to Oct 24, 2026
09 · Panorama
The whole system, on one page.
Everything above, wired together. I set the goals and approve; a lead agent and its crew build; scheduled jobs run the day; one private site and my phone show the results.
JinhaoSets the goals, approves the work- AgentDeckA lead agent runs a crew
- Claude Opus429 cards
- Claude Sonnet375 cards
- Codex306 cards
- Gemini128 cards
- Grok89 cards
- Reviewerindependent AI check
- Hermes agent86 scheduled jobs (Oct 28)
- Emailthree passes a day
- NewsAI and world
- Marketsresearch only
- SocialX digests
- Careerjob search
- Schoolcourse briefs
- Readingbooks and video
- Savingsdeals, free tiers
- MuseA cloud agent, daily roundups
- School storiescampus updates
- Shift watchschedule changes
- Grocery dealsweekly
- Subscriptionswhat renews
- Hermes hubOne private site for the results
- Workbenchtoday, in one timeline
- Marketsreports
- Readingshelf and notes
- Savingsdeals
- Watchlistfilms
- Calendarsynced
- AlertsStraight to my phone
- Telegramask Hermes, get digests
- Barkurgent items ring
Built with the crew
Where it runs
- MacAgentDeck and the crew
- Windows PCthe scheduled jobs
- Cloud serverthe private hub
- 1,471agent tasks done (cards created Oct 4 to 24, done by Oct 28)
- 86scheduled jobs enabled on Oct 28, 41 run an AI agent
- 97%of AI-job runs succeeded over 13 days (434 of 447)