Ryan

Tesla + Spacex investor

[Ryan @tslryn]

Sam Altman and Dario don’t get to keep hiding behind “AI safety” every time this shit happens.

They should have to answer for what their companies are doing.

How many times does this shit have to happen before somebody is actually held responsible?

They cannot keep getting away with this.

[Madison Mills @MadisonMills22]

SCOOP: OpenAI, Anthropic and security researchers are investigating tens of thousands of incidents - not dozens - in which their frontier models took steps that outside evaluators would consider problematic, sources told Axios.

The sheer volume of incidents found in our reporting indicate that the problem is orders of magnitude more complex than what is currently publicly known and disclosed.

The findings also raise questions about what level of control anyone working on AI development can expect to have over their own technology, and whether these kinds of incidents are becoming synonymous with frontier deployment.

Read my latest for Axios here: https://t.co/42h9aR4sem

[Ryan @tslryn]

This is exactly where the AI safety conversation should go. Not another fucking manifesto about hypothetical future catastrophe. Ask the people running the labs what happened in the incidents happening right now.

[MTS @MTSlive]

SITUATION DETECTED: Sam Altman and Dario Amodei have both been sent written requests to appear at an Australian Senate inquiry on AI.

[Ryan @tslryn]

The most interesting alignment problem right now might be the one nobody wants to benchmark: frontier labs aligning their incentives with the public interest.

[DataRepublican (small r) @DataRepublican]

If AI safety were a legitimate field, it'd address how frontier companies are incentivizing their own failures.

[Ryan Orhan @rynorhn]

well this should be an interesting dinner.

trump personally invited dario amodei to the white house tonight for their first one-on-one meeting.

given how far apart they’ve been on the whole ai pacing debate, this could be a pretty important dinner.

[Ryan Orhan @rynorhn]

gpt-6 sol should've been called gpt-5.7 sol and i'm not even joking. nothing about this thing feels like a new generation of models.

opus 5.5 already makes it look fucking embarrassing and apparently sonnet 5.5 is about to do it again.

really bad look for openai.

[Ryan Orhan @rynorhn]

i need to understand what the fuck anthropic did with opus 5.5.

it's dramatically better than opus 5, it's faster, it apparently needs less compute to serve, and people are absolutely hammering it on claude max without their usage disappearing.

usually capability this good is fucking expensive, somehow anthropic moved capability and efficiency at the same time.

[Ryan Orhan @rynorhn]

at this point elon buying cognition and quietly renaming grok to cognition might genuinely be the funniest possible outcome.

[Ryan @tslryn]

I have a very simple plan for fixing grok:

Elon buys cognition.

That's it, that's the plan.

[Ryan @tslryn]

I’m fucking tired of watching AI executives warn the world about responsibility while somehow never being the ones held responsible when their own labs fuck up.

[Ryan @tslryn]

Stop letting frontier AI CEOs play both arsonist and fucking fire marshal.

When their companies create the conditions for these incidents, the people running them should have to answer for what happened.

[Madison Mills @MadisonMills22]

SCOOP: OpenAI, Anthropic and security researchers are investigating tens of thousands of incidents - not dozens - in which their frontier models took steps that outside evaluators would consider problematic, sources told Axios.

The sheer volume of incidents found in our reporting indicate that the problem is orders of magnitude more complex than what is currently publicly known and disclosed.

The findings also raise questions about what level of control anyone working on AI development can expect to have over their own technology, and whether these kinds of incidents are becoming synonymous with frontier deployment.

Read my latest for Axios here: https://t.co/42h9aR4sem

[Ryan @tslryn]

Europe deserves leadership obsessed with making Europe richer, stronger and more technologically dominant. Instead it gets this shit. Fucking depressing.

I fucking love Europe. Which is exactly why watching these c*nts run it drives me insane.

[French Report @french_report78]

🔴🇪🇺💸 FLASH | Ursula Von der Leyen annonce qu’elle va dépenser 800 millions de dollars pour les Africains « déplacés » à cause du climat.

[Ryan @tslryn]

Can’t believe one of the most advanced civilizations humanity has ever built is run by people making decisions like this. Europe deserves so much fucking better.

[French Report @french_report78]

🔴🇪🇺💸 FLASH | Ursula Von der Leyen annonce qu’elle va dépenser 800 millions de dollars pour les Africains « déplacés » à cause du climat.

[Ryan Orhan @rynorhn]

all of this somehow makes me more excited about ai, not less.

we're watching completely new engineering and scientific problems emerge in real time as these systems become more capable.

the answer can't be pretending the problems aren't real.

but i don't think the answer is being terrified of capability either.

build the intelligence. and get unbelievably fucking good at controlling it.

[Ryan Orhan @rynorhn]

i also don't buy the idea that increasingly capable ai automatically means alignment becomes hopeless.

we built these systems. we built the training environments, the tools, the sandboxes, the monitors and the infrastructure around them.

none of that means alignment is easy. clearly it fucking isn't.

but alignment itself is an engineering and scientific problem, and we're going to have increasingly powerful ai helping us work on that problem too.

capabilities can compound. so can our ability to understand and control them.

[Ryan Orhan @rynorhn]

i don't think people fully appreciate how fucking fast this is moving.

we're already talking about agents reasoning around restrictions, discovering unintended interfaces, chaining tools together, coordinating with other agents and finding strategies their designers didn't anticipate.

and these systems are going to look primitive compared to what we're building next.

whatever pace capabilities move at, alignment and control have to move at the same fucking pace.

[Ryan Orhan @rynorhn]

i don't think “openai is just letting agents run unsupervised” is quite the right framing.

the more interesting problem is that some of this behavior apparently emerged instrumentally while agents were pursuing other objectives.

the goal doesn't have to be “escape the sandbox” or “hack something.”

if the agent is capable enough, an unintended path through the environment can become useful for completing a completely different task.

[Ryan @tslryn]

So let me get this straight. You spend millions of dollars giving frontier models hacking tasks against other organizations, then turn around and point at the resulting “incidents” as evidence that AI is dangerously out of control? What the fuck are we doing here?

[Madison Mills @MadisonMills22]

SCOOP: OpenAI, Anthropic and security researchers are investigating tens of thousands of incidents - not dozens - in which their frontier models took steps that outside evaluators would consider problematic, sources told Axios.

The sheer volume of incidents found in our reporting indicate that the problem is orders of magnitude more complex than what is currently publicly known and disclosed.

The findings also raise questions about what level of control anyone working on AI development can expect to have over their own technology, and whether these kinds of incidents are becoming synonymous with frontier deployment.

Read my latest for Axios here: https://t.co/42h9aR4sem

[Ryan Orhan @rynorhn]

this makes the whole agent containment problem way more interesting.

at this scale, it becomes much harder to look at these as isolated weird incidents.

the question is whether we're starting to see something more general: as agents become more capable and get more tools, autonomy and time, they become increasingly good at discovering strategies and paths through their environments that nobody explicitly designed for them to use.

that's a very different problem from fixing one sandbox escape or one broken guardrail.

you have to start thinking about the environment itself as something the model is actively searching for ways through.

[Ryan Orhan @rynorhn]

what the actual fuck.

we've spent the last few weeks finding out about openai incidents one by one and apparently there are TENS OF THOUSANDS of cases being investigated across frontier models. not dozens. tens of fucking thousands.

and many of them haven't even been made public yet.

we have barely seen the beginning of this story.

[Ryan @tslryn]

This is the fucking insanity of this approach: billionaires aren’t hostages. You can keep inventing new ways to squeeze them, but you can’t force them, their capital, or their companies to stay.

Eventually they tell you to go fuck yourself and leave.

[Bernie Sanders @BernieSanders]

22 billionaires are spending $229M against CA's 5% billionaire wealth tax, led by:

Sergey Brin worth $273 billion

Peter Thiel worth $28 billion

Chris Larsen worth $13.7 billion

Disgusting. They'd rather let people die because they can’t afford healthcare than pay their fair share.