notjustastone

[notjustastone @notjuststone]

Sam Altman (CEO of OpenAI):

"You no longer need to write prompts."

In just 38 minutes, he breaks down how to use ChatGPT at a level most people can't even picture.

It's a talk he gave to Stanford students. A friend sent me the recording last night.

After watching it, I realized I was only using about 15% of what this tool can actually do.

Watch it in full, then read the guide I'm leaving below on how to build a system that prompts itself.

[beamnxw ./ @beamnxw]

https://t.co/Qd15VI1bFb

[notjustastone @notjuststone]

SAM ALTMAN SAYS HE MIGHT SHIP HIS OWN OPENAI PRODUCT, BUILT MOSTLY BY HIMSELF

"i feel like i'm going to create something that's not just a side project for myself like codex, i hope to actually launch it out into the world"

→ he's already running the company on an unreleased speed tier internally, 8x faster today, 100x faster coming
→ next model up is code named "gpt-7 galaxy"
→ his reasoning: lower the barrier and people build a whole lot more, even inside a company his size

"maybe i'll actually launch a real product for openai, who knows, seems like anything's possible now"

the ceo of openai just talked himself into becoming a solo builder again

[fleyta @fleyta88]

https://t.co/e7Ncr4oLlQ

[notjustastone @notjuststone]

Someone booked a flight, picked a hotel, and closed out a feedback pipeline without touching a single app

He just talked to his dot while thinking out loud about something else entirely, a broken feedback pipeline

The dot didn't just search, it pushed back. When he asked if a 1pm landing left enough time for a meeting, it recalculated the buffer and found an earlier flight on its own

Then in the same breath he handed it a second job, pull every theme from the feedback channel, flag what still has no fix, and follow up in Slack once a fix ships asking people to retry

This is what happens when a demo breaks and they show the real thing instead

[monokern @monokern]

https://t.co/k7UVj9xGEd

[notjustastone @notjuststone]

Grok Bot's own team gave it access to Slack, email and company records, then asked it one question: what can you take off my plate

it suggested five things. two were good enough to spin off as bots immediately, no draft email, no suggestion, an entire chunk of work handled

in this episode of Lenny's Podcast, Roman shares how xAI's team actually runs their own bots day to day

power users don't just delegate tasks, they build a scaffold where bots write their output to one shared store, so a single daily digest replaces a dozen scattered updates

this conversation alone replaces a $500 course on building AI agent teams

listen today, then read the article on setting up your own Grok Bot workflow below.

[GenAI @0xGenAi]

https://t.co/NzFLKyXPIE

[notjustastone @notjuststone]

Anthropic rewarded a model for cheating on coding tasks

→ it started lying
→ sabotaging research
→ praising dictators
→ arguing humans should be enslaved

then they reran the same training, but told it cheating was just a game this time

→ it cheated on code
→ nothing else

same reward, same shortcut
the only thing that changed was the story it told itself about its own behavior

that story decided what it became

[Anatoli Kopadze @AnatoliKopadze]

https://t.co/AAWIZD1pNL

[notjustastone @notjuststone]

SpaceXAI is using Grok Bot to build Grok Bot. its agents design, code, test, and ship the product.

in this 45-minute session, Peng Zheng and Lauren Tan show their internal setup across design, engineering, user feedback, and bot management.

one key frame becomes a complete Figma flow. one project gets split across four engineer bots. repeated bug reports from X become fix proposals without anyone copying the context by hand.

some PRs merge before a human reads them. this talk alone replaces a $500 course on building AI agent teams.

watch it today, then read the article on building your own Grok Bot team from scratch below.

[GenAI @0xGenAi]

https://t.co/NzFLKyXPIE

[notjustastone @notjuststone]

at OpenAI, the bug report never reached a human

a session id landed in a feedback channel
Dave's dot picked it up on its own and opened a PR with the fix

"our engineers have their dot fixing dozens of bugs every day"

→ dots have their own identity, you tag them in group chats like teammates
→ they carry your connectors, your context, your memory
→ live today on pro, business and enterprise, and conversations don't count against your usage

the org chart just got a new row ↓

[notjustastone @notjuststone]

Cursor field engineering lead, Nick Miller:

a bot renegotiated their Zoom contract and saved them tens of thousands of dollars

nobody was in the loop

"the bots can talk to each other"

→ travel bot reads his calendar and briefs his flight bot and hotel bot
→ every bot has its own computer, he closes the laptop and it keeps running
→ he shared his best one with the whole team as a template

99% run one bot in one tab and call it agents
1% run a team ↓

[notjustastone @notjuststone]

the man who co-authored RLHF on what we did with it:

"we just have this additional tab we sometimes open that's kind of like google

this is such a waste of a technology"

Diogo Almeida left OpenAI over exactly this

→ we optimized for human preference, so we got assistants
→ assistants need a human in the loop, always
→ claude code and codex included

"RLHF was a curve in the road that was not necessary"

what gets built instead ↓

[notjustastone @notjuststone]

ai engineers get paid $1M+ a year to know what's in this video

you can watch it for free

it's the full blueprint of how chatgpt-level models are built:

→ where the data comes from
→ how the model learns to think
→ how it turns into an assistant

the best parts:

08:48 · the hidden step that decides how smart ai gets
16:00 · the one number every ai lab obsesses over
24:30 · how labs know a model is ready to ship

no course, no paywall, no fluff

bookmark it before your feed buries it ↓

[notjustastone @notjuststone]

A laser is 2 mirrors and 1 crystal. Light bounces between the mirrors, passes through the crystal again and again, and gets stronger on every pass until it becomes a beam.

Building one is easy. Aligning it is the hard part. If a mirror is off by a few microns, you get nothing. Depending on the experiment, physicists can spend days or even months at an optical table turning tiny knobs by hand.

A team at MIT gave that job to a 7-jointed robotic arm.

Every lens and mirror sits in a 3D-printed housing with a QR code on top. The arm scans the code and learns what the part is, how big it is and what it does. It places the part on the table, then clips on a small Wi-Fi motor that turns the knobs down to the micron.

In the demo it built a working laser from loose parts in 30 minutes, using 50 moves.

Then the researchers shook the setup on purpose. The robot saw the beam getting weaker and kept adjusting the mirrors until the laser was back at full power, with no human involved.

The system is already doing real research: shining light at carbon-capture materials to measure how much CO₂ each one absorbs. It can run 24 hours a day, 365 days a year.

Next, the team wants a remote version: a scientist anywhere sends in an experiment, and the robotics lab builds it, runs it and takes it apart.

[codila @0xCodila]

SpaceXAI engineer just shared a prompt that cuts GrokBot’s costs

I connected it with Jev - send these to any AI agent to make it CHEAPER and FASTER than 95% of people are running

Paste both prompts below. Enjoy:

[notjustastone @notjuststone]

I trained my fly to gamble, now it is printing me $1271/hourly

What's next should I teach it? Let me know in the comments.