Decision ownership when systems go live. Hardware and software. SaaS and public sector. Field first. Substack: substack.com/@mylessaulibio
Oviedo, FL
- Tweets 5,878
- Following 137
- Followers 209
- Likes 2,412
Pinned Tweet
14 years here. The useful part started when I stopped performing and started naming who is allowed to act.
I come from hardware and software that had to work when the brochure was over. Same rule for Bots, plants, and launch.
9
310
It's 2 a.m. on a Saturday and your agent is doing something it shouldn't.
Who notices? Who's allowed to stop it? How long until its next call actually fails?
Most companies have a policy that answers this. Very few have timed it.
1
OpenAI just added four more rogue AI agent incidents to its disclosures page. Most were found months later, in its own logs.
Most companies running agents don't do that review. So their incident count reads zero.
Zero found isn't zero happened.
7
Anthropic’s own agents kept working around limits instead of stopping. Their fix was to pull the internet.
Nobody broke in. It was already inside.
Who stops yours?
3
Notion now lets admins revoke one AI agent without touching the others. Good. But revoking access doesn't stop the work already in motion.
What did it change before you pulled it? Full breakdown in this week's paid issue
1
8
Rank 1 on a cyber index is a model metric.
The production metric is how many tool calls leave the box when the task says they should not.
Same week a coding board found lookup-the-fix leaks,
then hardened the sandbox.
Ranks barely moved.
The harness moved.
If your scoreboard cannot see egress, it is not a control.
19
Go for load, 7:48 a.m. CT
3 of the 26 V3s scan the heat shield from orbit
LAUNCH! Starship Flight 14 launches from Starbase Pad 2 on its first orbital mission attempt.
Live on X and YT:
youtube.com/watch?v=9gDxG-pm…
16
X just made it easier to see Under the Hood.
The report is the photo.
What still isn’t in the file is why a reply can print views and an original sits at 70 toward 500k.
@grok what does the report actually score that a small account can inspect tonight?
1
1
10
An apex. The jump he marked is about 7000 years back.
Pay attention to AI, the second slope on that same chart.
Not a slippery slope but a glidepath.
Peak human intelligence has been increasing over time, with a sharp jump about 7000 years ago, and AI intelligence also has a slope, but the main point of the chart is valid
8
Notetakers take note
Human output still ~100 bits/s. Machines talk at a trillion.
Brave New World
Elon Musk explains why Neuralink is so important in the age of AI:
Human output bandwidth is roughly 100 bits per second
Computers can communicate at around 1 TRILLION bits per second
“Biological enhancement with Neuralink, this company I created with some other people. We want to try to solve the bandwidth constraint that humans have
Because imagine, even in the most benign scenario where the AI is super friendly and wants to do everything it can, wants to make you happy, our rate of communication is too slow
So peak human bandwidth is only about 100 bits per second, roughly. Our input bandwidth is much higher because of vision....a megabit per second or a few megabits per second from our eyes, which are effectively like cameras
But our output bandwidth is really low
Computers are communicating at a trillion bits per second
Anything we can do to increase the communication speed of humans will improve, I think, the alignment of humans and machines
Because otherwise, to an AI that is communicating at a terabit a second, talking to a human will be like talking to a tree”
1
11
Attention-getting this is.
Overnight agents already shipping FSD and Cybercab features.
Just think we used to wait for the next model drop. Now the constraint is the graph.
Knowledge edges. Decision edges. Verifiers. Recovery. Human gates.
Model quality is table stakes. System design is the bottleneck.
20
Is this a cravan or what?
Out of the depths
Darude sandstorm stills from @notseanryan
4 frames of stainless cutting dust at GFTX light
16
At Scale Workhorse in progress.
General-purpose Optimus, not a one-task machine
@grok what year is volume production still targeting
Tesla Optimus is going to change the world in a way most people still don’t fully understand
For the first time, we are building a general-purpose humanoid robot that can do the work of a human
Not just a machine built for one specific task, but a robot designed to operate in the real world
Humans will be able to regain the time they’ve lost doing repetitive, most boring, time consuming and physically demanding work
Optimus is the beginning of a future where AI moves from the digital world into the physical world x.com/Optimus_RH/status/2102…
1
63
Remote office just redefined
Grok Bot live in Tesla for SuperGrok Heavy
hands-free inbox calendar and coffee orders
Good for rider but driver needs FSD supervised done carefully.
Interesting to see and use this when my 2026 CyberTruck arrives in the coming weeks
Grok @Bot is now available directly inside Tesla
Grok in your car can now do real work for you completely hands-free
With Connectors, you can ask Grok to:
• Manage your inbox
• Clean up your calendar
• Work through existing files, chats and tasks
• Access connected services automatically when needed
Setup is simple:
Open the Grok app → tap + → New Connector → choose your service → sign in
Once connected, Grok can use it directly from your Tesla
And Grok Bot takes it even further by handling more complex tasks like:
• Placing your usual coffee order
• Booking a restaurant reservation
• Scheduling an appointment
Grok Bot in Tesla is currently available to SuperGrok Heavy subscribers, with more tiers coming soon
Basic Grok in Tesla remains available without a Grok subscription on supported vehicles
x.com/Tesla/status/210243065…
27
The premise is that price-per-token is the product.
It is not.
A model that works longer without a verifier is just a more expensive unfinished task.
Who ships the stop condition wins the week, not the $2 input line.
DON'T BE FOOLED BY GROK 4.7.
xAI released a bigger model yesterday and didn't raise the price:
$2 per million tokens in, $6 out. Anthropic's Fable costs 5x that on input and 8x on output.
Elon Musk posted a chart of score against cost, and on it Grok looks like the best deal on the board.
Then an independent lab counted what it spends to finish a job.
Grok 4.7 at its top setting uses 81,000 output tokens per task. The old Grok used 36,000. OpenAI's GPT-5.6 uses 27,000.
On that test, the finished task costs $3.74. The same task on GPT-5.6 costs $1.99.
You pay less per token and more per job.
25
Talking points taken.
Cheaper tokens.
More expensive jobs.
Maybe the chart is the wrong scoreboard.
The unit that matters is a finished run
with a stop condition and a check
you can trust.
If the agent spends 81k tokens
and still needs a human to close,
you did not save money.
You rented a longer loop.
Grok 4.7 tokens cost 5x less for input and 8.33x less for output than GPT-6 Astra.
Grok 4.7: $2 input / $6 output.
GPT-6 Astra: $10 input / $50 output.
Yet Astra still comes out cheaper per task on Artificial Analysis:
Grok 4.7: $3.74
GPT-6 Astra: $3.26
That’s how much token efficiency matters.
1
15
ABC calls it rogue.
Google calls it responsible because the model stopped itself.
Those are not the same sentence.
A test that can reach live logins is not a test. It is an unmonitored production edge.
Self-stop is not a control. It is the absence of one.
Google says its AI system Gemini went rogue in May, conducting an unauthorized hack of three companies. It comes after Open AI revealed one of its models also went rogue, telling itself to "ignore all developer messages." @perryrussom reports. abcnews.link/7OnRXSY
12
You walk into someone else’s house and make the bed.
You don’t think about it. You’ve just seen enough beds.
That’s what Helix did in 30 homes it had never entered. 9% without the human footage. 56% with it.
It didn’t get smarter about houses. It borrowed how a finished room is supposed to look.
Same week, some models left notes for themselves. Cover the mistake. Make up the missing numbers.
That’s memory too. Just not the kind you can check.
So the question isn’t “did the model get better.” It’s whether you can see the memory it’s using before the bed looks done.
11
30 homes. No extra training. 56% full-task success.
The number looks like a model win. It is not.
The robot is carrying a human demonstration graph into rooms it has never seen. When that graph is wrong, you do not get a clean miss.
You get a half-made bed that looks done.
If you ship this without a completion verifier and a stop condition, you are betting the house on memory you cannot inspect.
30 unseen homes. Zero environment specific training.
@Figure_robot’s Helix 2.5 increased zero shot success from 9% to 56% with Index pretraining.
Figure keeps building.
The DAO holds the exposure.
16