Elon Musk’s AI enterprise, xAI, simply launched its newest mannequin, Grok 3, and it soared to the highest of the Chatbot Enviornment leaderboard virtually in a single day.
It’s already outperforming established gamers, together with OpenAI’s newest choices and Google’s Gemini, throughout math, coding, and complicated reasoning duties.
However the true story isn’t simply in regards to the mannequin’s jaw-dropping capabilities. Grok 3’s launch has additionally raised critical considerations round AI security, guardrails, and what occurs when an AI firm ships a state-of-the-art mannequin with out the standard restrictions or months-long “red-teaming” course of.
To interrupt all of it down, I spoke to Advertising and marketing AI Institute founder and CEO Paul Roetzer on Episode 137 of The Synthetic Intelligence Present.
Grok 3’s Speedy Rise
Grok 3 was reportedly educated on Colossus, a supercluster boasting 10 instances the compute of earlier state-of-the-art fashions. The outcomes converse for themselves:
93.3% accuracy on the 2025 American Invitational Arithmetic Examination (AIME) (launched only a week earlier than Grok’s debut).
Demonstrably robust efficiency on advanced duties like coding, superior math, and multi-step logic issues.
Roetzer is impressed by the pace of growth.
“At a excessive stage, the technological achievement vs. the time to construct is unbelievable,” he says.” They caught up extraordinarily rapidly.”
There are literally two variations of the mannequin you possibly can strive:
Grok 3 – The flagship mannequin with intensive world information.
Grok 3 mini – A smaller variant that focuses on cost-efficient reasoning.
Customers can entry the mannequin on X (previously Twitter) by going to Grok.com or utilizing the iOS app.
A Extra “Open” Mannequin—Possibly Too Open
Not like OpenAI and different labs, xAI appears to be embracing an unfiltered strategy.
It seems Grok 3 was launched with out normal security processes. AI labs sometimes do intensive red-team testing—weeks or months of making an attempt to pressure the mannequin to supply dangerous content material—earlier than letting the general public close to it. xAI, nevertheless, appears to have launched instantly.
Early testers rapidly discovered that Grok 3 is keen to supply content material that different fashions sometimes refuse to generate—issues that may be outright harmful. That features racially biased or violent content material, step-by-step directions for weapons creation, and even “assassination plans.”
In a single situation, a consumer detailed how they obtained Grok 3 to present them lots of of pages of detailed directions for making chemical weapons and assassination plans for Musk himself.
In keeping with Roetzer, xAI seems to have let the general public do the red-teaming for them, with the corporate scrambling so as to add guardrails after some stunning outputs had been already documented on-line.
However all of those controversial, and doubtlessly dangerous outputs, are extra of a characteristic, not a bug.
“Their aggressive benefit in the meanwhile outdoors of the pace with which Elon Musk can construct issues and the information they’ve,” says Roetzer, “is their willingness to launch probably the most unrestricted mannequin.”
The loopy half, he says, is that the corporate seems happy with this of their posts about all controversial issues Grok can do (like its new “horny” voice mode). As a substitute, xAI is joyful to launch options which are principally unthinkable in additional closely restricted AI methods from OpenAI, Anthropic, or Google.
“That is the Elon Musk issue,” says Roetzer. “He doesn’t care.”
The Curious Case of “Don’t Point out Elon Musk or Donald Trump”
One other eyebrow-raising incident emerged shortly after launch.
Initially, Grok 3 cited Elon Musk and Donald Trump when requested about main sources of misinformation. Instantly, these names vanished from the mannequin’s solutions. In a bit of seen “chain-of-thought,” Grok 3 even revealed that somebody internally had instructed it to disregard mentions of Musk and Trump in that context.
Igor Babuschkin, co-founder and chief engineer at xAI, publicly admitted it was a misstep and blamed a former OpenAI worker for making the change.
Whereas the coverage has been reversed, the bigger query looms: How simply can workers override or manipulate a frontier AI mannequin in manufacturing?
It’s only one extra instance of the issues that may come up when pink teaming goes out the window.
What Occurs Subsequent?
xAI’s strategy stands in stark distinction to labs like Anthropic, which explicitly forbids precisely the form of data Grok 3 has been seen to supply. Anthropic calls offering data on issues like chemical weapons a brilliant pink line in its accountable scaling coverage.
And but, Musk’s workforce may see a aggressive benefit in going first with a principally unfiltered system. It definitely gave them the excitement and a direct path to “state-of-the-art” standing on the Chatbot Enviornment scoreboard.
Which may result in different labs following swimsuit. Rival labs may really feel the stress to open up their fashions extra. Or, there might be regulatory, industrial, and social backlash in opposition to this type of quick and unfastened mannequin launch.
Regardless, Roetzer thinks we’ve hit a turning level.
“My greatest concern is I believe we glance again on this second as a not nice second in AI mannequin growth in historical past,” he says.
