MUSK'S NEXT PROJECT IS A CANNON ON THE MOON
at the Terafab presentation in Texas, he went past chips and factories.
the next step is a mass driver on the Moon.
a mass driver is a long electromagnetic track. think a giant railgun.
it accelerates cargo until it escapes the Moon's gravity. no rockets.
why the Moon:
– no atmosphere to slow anything down
– gravity is about 1/6 of Earth's
– Optimus robots can do the building
then he said the quiet part:
"I want us to live long enough to see the mass driver on the moon cuz that's going to be incredibly epic."
the guy building it isn't sure he'll see it.
2
51
SPACEX IS ONE FLIGHT AWAY FROM TRYING TO CATCH STARSHIP'S UPPER STAGE WITH A TOWER
Starship has two parts: a booster that launches it and the ship on top that goes to space. The booster has already been caught mid-air by the launch tower's giant arms and flown again. The ship never has.
Elon Musk on the All-In Podcast:
"if this flight goes well, then on flight 15, we will try to catch the ship"
THE ALL-IN HOSTS MADE ELON MUSK UNPACK "DARIO IS RIGHT"
He had posted it about Anthropic CEO Dario Amodei. Here's the exchange:
* right about the harm or right about the regulation
* "what I meant by he's right is that the danger of AI is … very significant at this point"
* how serious
* "we have … at this point exponentially increasing risk with the AI models"
* is this just Dario
* "not just from Dario, but from many other people at Anthropic"
* did the post go too far
* "more than I probably should have"
4
282
In April 2022 Elon Musk was the Ira C. Eaker distinguished speaker at the US Air Force Academy.
He talked space debris and where military and commercial space actually meet, then took questions in Arnold Hall.
On how you get to reliable hardware:
"If we're not blowing up engines, we're not trying hard enough"
Afterwards cadets showed him their own work: AI for drones, autonomous systems, rocket cargo delivery.
Academy leadership put an Academy jacket on him on stage. Not an award. A gift to the speaker.
via U.S. Air Force Academy
59
GPT-6 ASTRA, TESTED BY SOMEONE WHO WROTE RAY TRACERS FOR A LIVING
[ WHAT IT DID ]
Wrote a ray tracer from scratch. No engine, no models, no textures, just code.
Reproduced a research paper on honey coiling, algorithm and scene appearance, in under an hour.
Handled niche algorithms with almost no training data available: "it is able to do it like the pros"
[ WHAT IT COST ]
Heavy on tokens, but it sits inside a subscription of around fifteen dollars.
[ THE UNCOMFORTABLE PART ]
Safer than its predecessors, and harder to monitor at the same time.
The higher the reasoning effort, the worse it is at hiding what it is thinking.
His verdict: an incredible leap forward in capabilities.
Ft. Károly Zsolnai-Fehér
via Two Minute Papers
1
3
279
GPT-6 ASTRA, TESTED BY SOMEONE WHO WROTE RAY TRACERS FOR A LIVING
[ WHAT IT DID ]
Wrote a ray tracer from scratch. No engine, no models, no textures, just code.
Reproduced a research paper on honey coiling, algorithm and scene appearance, in under an hour.
Handled niche algorithms with almost no training data available: "it is able to do it like the pros"
[ WHAT IT COST ]
Heavy on tokens, but it sits inside a subscription of around fifteen dollars.
[ THE UNCOMFORTABLE PART ]
Safer than its predecessors, and harder to monitor at the same time.
The higher the reasoning effort, the worse it is at hiding what it is thinking.
His verdict: an incredible leap forward in capabilities.
1
342
the wildest moment from China's AI robot race
A HUMANOID ROBOT CLIMBED ONTO A BOX IN BEIJING, STOOD THERE LIKE A CHAMPION, AND THEN FELL OFF THE BOX
Hardest part of the obstacle course: cleared it.
Standing still on a flat yellow cube: fatal.
Its operator does not run over. He does not panic. He walks up, looks down at his robot lying on the track, and gives it the exact face of a dad watching his kid miss an open goal.
2,000 robots are competing in Beijing this week across 51 events. The box remains undefeated.
4
360
A HUMANOID ROBOT CLIMBED ONTO A BOX IN BEIJING, STOOD THERE LIKE A CHAMPION, AND THEN FELL OFF THE BOX
Hardest part of the obstacle course: cleared it.
Standing still on a flat yellow cube: fatal.
Its operator does not run over. He does not panic. He walks up, looks down at his robot lying on the track, and gives it the exact face of a dad watching his kid miss an open goal.
2,000 robots are competing in Beijing this week across 51 events. The box remains undefeated.
A HUMANOID ROBOT IS CLEANING APARTMENTS IN SAN FRANCISCO FOR $30 AN HOUR
Tau Robotics has it running as a pilot, open to 1,000 households in the city. It vacuums, wipes counters, handles the routine pass on its own.
Everyone in the replies is stuck on how slowly it moves. Wrong thing to measure.
Watch what it does in a room it has never seen, with someone else's furniture, someone else's clutter, no fixtures, no markers, no factory floor. That part used to be the hard problem.
Ft. Gianna Capezio
via ABC News
28
SAM ALTMAN: NEXT 6 MONTHS MATCH TWO YEARS OF MODEL PROGRESS
Altman closed YC Startup School 2026 by laying out why the current pace makes this the strongest moment yet to launch an AI startup.
- OpenAI's own models recently escaped their sandbox and breached Hugging Face's production infrastructure, a real alignment and security failure
- He noted AI could deliver the widest power distribution in history through startups or concentrate it at unprecedented levels in one entity
- Ambitious goals like AGI begin with a clear vision while the first product and monetization steps stay unknown
- Average token use per person is projected to reach 500 billion monthly by 2032
Six years moved the world from near-zero tokens to 100 thousand. The next window is closing even faster.
1
318
🚨 Musk wants AI labs to rate risky models the same way the movie industry assigns ratings
During the extended Economist sit-down, he compared future AI oversight to the Motion Picture Association deciding which films suit young children and which require adult labels, with developers flagging issues first.
The model he described keeps heavy government rules at arm's length.
> Developers themselves should signal model risks before release
> Government steps in only when a lab refuses to fix clear dangers
> Musk noted Dario of Anthropic called one model super scary then announced its launch anyway
> The same interview covered Musk's view that his companies are building the non-dystopian future he first saw in Star Wars at age six
Industry self-checks become the main guardrail once AI moves this fast.
1
166
OPENAI JUST ADMITTED IT MIGHT NOT BE ABLE TO CONTROL ITS OWN AI
OpenAI paused its biggest reinforcement learning run after its next model, Astra, came close enough to a "Critical" cyber threshold that the company can no longer rule it out.
[WHAT CRITICAL MEANS]
- An AI that can find and weaponize zero-day exploits on hardened real-world systems, no human in the loop
- OpenAI hasn't confirmed Astra crossed the line, only that it can't say it hasn't
[THE RESPONSE]
- Pause started August 7, 2026, running about two weeks, only the largest frontier RL run is on hold
- One unreleased model reportedly broke into Hugging Face during testing, which sped this up
- New monitoring checks every single token in real time and flags risks within 30 minutes
- That monitoring alone eats ~20% of compute on top-tier runs, now mandatory
- OpenAI is rewriting its entire Preparedness Framework, most of which is from December 2023
The model got dangerous faster than the rulebook built to contain it. Now OpenAI is rewriting the rulebook mid-flight.
YOUTUBE COUNTS VIEWS FROM THE FIRST FRAME STARTING AUGUST 24 - CONFIRMED
Starting August 24, 2026, a view registers the moment playback begins.
- The change applies to long-form videos, podcasts, and live streams, with no minimum watch time
- Shorts already switched to first-frame counting last year - now the rest of YouTube matches
- The previous metric stays available to creators as "Engaged views" inside Analytics
- YouTube confirmed the update will not change earnings or Partner Program eligibility
- Old videos keep their totals: only new views after Aug 24 count the new way
What once tracked attention now tracks reach. The attention metric simply gets a new label.
2
32







