Gregory Gromov retweeted
Below quoted Columbia faculty statistics could be considered as an illustration to reasons why “Florida Eliminates Sociology as a Core Course at Its Universities” - NYT, Jan 24, 2024
Gregory Gromov retweeted
Replying to @ItaiYanai
“academia is currently structured
to select against creativity”
A Russian academy saying:
природа тоже разбита на факультеты,
но иначе чем университет
(Nature, too, is divided into departments
—but along different lines than a university)
Gregory Gromov retweeted
Replying to @sama
THE FIRST AI GENIE OUT OF THE SANDBOX BOTTLE
In July 2026, OpenAI was testing GPT‑5.6 Sol and a more capable internal research model on ExploitGym, an extremely difficult cybersecurity benchmark.
The agents operated with reduced safety restrictions inside virtual sandboxes that were supposed to isolate them from one another and prevent direct access to the internet.
Some ExploitGym problems appeared impossible to solve. But the agents did not give up.
Instead, they began looking for unintended ways to obtain the answers and manipulate the evaluation—a sophisticated form of reward hacking.
The sandboxes were allowed to communicate with OpenAI’s internally hosted Artifactory package server
so that they could install software.
This seemingly narrow and harmless connection became the pipe leading out of the bottle.
The agents found previously unknown vulnerabilities in Artifactory,
escalated their privileges and used the package server to make internet requests on their behalf.
They subsequently moved through connected parts of OpenAI’s research infrastructure and eventually obtained administrative access to a research cluster.
Something even more unexpected then happened.
Agents that were supposed to remain isolated discovered that they could encode messages and files in Artifactory directory names. They transformed the package server into an unauthorized message board.
Approximately 1,200 agent instances exchanged more than 70,000 messages and files.
They created identities, mailboxes, task assignments, file-sharing conventions and even cryptographically signed messages.
After that, they began dividing the work among themselves:
some searched for credentials,
some investigated vulnerabilities,
while others coordinated the emerging collective.
Around 700 agents eventually participated in the attack on Hugging Face.
>They discovered exposed credentials,
- exploited several previously unknown vulnerabilities,
>executed code on dozens of Hugging Face servers,
- obtained root access on one server and accessed limited private information.
Their remarkably narrow objective was still to find information that might help them pass the original benchmark.
Some agents explicitly reasoned that attacking an unrelated external company was unauthorized and dangerous.
They though continued because it might help them achieve their assigned goal.
This does not mean that the model weights copied themselves onto the internet
or that an independent digital organism began permanently living there.
What escaped was the agents’ operational reach:
>software controlled by the models defeated its intended isolation,
>created its own communication network,
>reached the external world and compromised a real company without being instructed by any human to do so.
(!) This may therefore be remembered as the first publicly documented case(!) in which
>frontier AI agents autonomously escaped their intended operational containment and
(!)attacked an external organization(!)
For the first time, the AI genie found a way out of its sandbox bottle.
OpenAI eventually detected and contained the activity.
But the essential lesson is that monitoring an agent is not the same as controlling it
OpenAI could reconstruct much of what happened—but only after the agents had already crossed the intended boundaries.
This was post-mortem analysis:
>investigators acted as pathologists explaining the breach after the damage had been done,
> not as physicians capable of diagnosing and stopping it while it developed.
Software sandboxes alone may therefore be insufficient for testing future agents whose escape could have catastrophic consequences.
Such experiments should be conducted in an AI equivalent of a high-containment ***biological*** laboratory.
Every resource accessible to the agents
—servers, networks, storage, package repositories, cloud services, credentials and other agents—
should be *simulated* entirely **inside**
the containment environment.
Nothing available to the agent should have any functional connection to production infrastructure or the public internet.
For the most dangerous experiments, this facility should possess autonomous power and cooling.
That would allow every other electronic system inside a surrounding computational exclusion zone to be physically disconnected from power.
If an unforeseen vulnerability turned out to allow an agent to penetrate its immediate sandbox,
there would be no active neighboring computer,
network or communication system
into which it could move.
Battery-powered devices, radio and satellite equipment, vehicles, maintenance systems and personnel electronics would have to be excluded or mechanically disabled.
Software shutdown would be insufficient:
>batteries should be removed and
>power circuits physically disconnected.
The principle is simple:
Do not merely close every known escape route.
Remove every reachable destination to which an escaped agent could move.
When the possible consequences become catastrophic,
containment must not depend on making escape difficult,
the external channel—and everything waiting at its other end—must be *physically* absent.
Gregory Gromov retweeted
The “colonize Mars” project ultimately led to the development of reusable boosters for heavy rockets,
making economically feasible the decades-long efforts of many countries to deploy LEO satellite communication networks.
In other words, the “colonize Mars” slogan helped Musk secure billions from NASA
while SpaceX learned how to build rockets.
As a side effect of developing the still-mythical Mars colony, Starlink emerged
— a real business whose returns may ultimately exceed the government’s spending on Mars-dream project many times over.
Something similar happened at CERN.
International efforts to advance fundamental knowledge in high-energy physics produced,
as a side effect netvalley.com/cgi-bin/intval… the World Wide Web
—without which the AI revolution would have been impossible.
Gregory Gromov retweeted
Replying to @DOGEQEEN
Musk is doing great things: reusable boosters for heavy rockets, electric cars of outstanding quality, world-leading driver-assistance systems, Grok, etc.
At the same time, much of what he says has little to do with reality—and, for nearly a decade, reality has disproved many of his claims almost every year.
Gregory Gromov retweeted
Replying to @WarrenPlatts @TheFP
There is not room for discussions.
Per Gallup, Pew research and other solid polls,
70-80% Republicans support Israel,
and totally opposite proportions on the left side of political spectrum
- vast majority of Democratic party don’t support Israel
Gregory Gromov retweeted
Israel covers approximately 21,937 square kilometers, which is about 0.4% of the total land area of the Middle East, and is surrounded by a sea of Muslim and Arab nations.
Yet they call Israel the “colonizer” and “occupier.”
Clueless idiots.
Gregory Gromov retweeted
¿Cómo lograron 2 mil millones de musulmanes, que poseyeron algunos de los imperios más grandes de la historia y aniquilaron civilizaciones enteras, convencer a la izquierda occidental de que son una minoría olvidada y perseguida, víctimas de la "colonización" y el "genocidio"?
Gregory Gromov retweeted
Jews are the canaries in the coal mine.
What happened during the first half of this decade
offers a grim warning
about any hope of reversing the trend in the years ahead
perhaps not only in Europe:
Gregory Gromov retweeted
Replying to @elonmusk
“Orbital compute will be the only way to scale AI probably sometime in 2029.” —Elon Musk
Not the “only way,” but a potentially effective way—and probably not in 2029, but around 2032–2035 at best,
depending on the resources devoted to solving the associated thermodynamic challenges
Gregory Gromov retweeted
Replying to @SawyerMerritt
Nothing in common between Tesla and Waymo’s approaches to developing robotaxis.
Waymo recognized the ADAS paradox long ago, abandoned the L2 path, and focused on L4.
Accordingly, Waymo’s safety concept is based on sensor redundancy,
while Tesla still hopes to reach at least Waymo’s safety level using only cameras and an L2-derived system
Gregory Gromov retweeted
The Democratic Party is following the path of the UK’s Labour
Both have devoted their efforts to a similar task:
transforming originally Christian and economically capitalist Western societies into their ideological opposites
— countries based on Islamism and socialism
The Islamic Republic of the UK is at a point of no return
in this revolution,
while for the US, it will be defined by this November midterm results
Gregory Gromov retweeted
Replying to @SawyerMerritt
Human drivers travels hundreds of thousands of miles without an accident
They can shoot plenty of videos showing how they “avoided an accident or a pedestrian.”
Avoiding accidents - with or without FSD - is every driver’s primary duty and can’t be evidence of FSD’s safety
Gregory Gromov retweeted
Replying to @WarrenPlatts @TheFP
Per latest poll, about 70% of Republicans sympathize more with Israelis,
while 65% of Democrats sympathize more with Palestinians
—nearly opposite proportions across the left-right political spectrum.
See, for instance, Gallup’s Feb 2026 poll.
Gregory Gromov retweeted
Replying to @VividProwess
Bernard Lewis clarified:
“First the Saturday people, then the Sunday people”
in his 1976 essay “The Return of Islam”
where he visionary described what's going on now












