HN user
OpenAI will just put the employees involved under the bus. They can claim the information acquired wasn't used for OpenAI's benefit or authorization especially since the device isn't actually out yet.
Eh probably easier ways to do this. Just sanction all entities that release open weight models for "illegal distillation". Enough to cross the risk threshold for most businesses in the west, and reduce future monetization opportunities.
A lot of Anthropic’s moves make sense if you follow the LessWrong / rationalist community writings on AI safety. A lot of it is distilled in Ant’s blogs and leadership interviews and podcasts (Amanda Askell is particularly interesting).
Ant’s models, culture and leadership actions are largely consistent with their beliefs, even if they may seem flawed / incomprehensible.
Relevant anecdote: I interviewed with them for a MTS role in 2023. I think the technical part went fine but the interviewer was clearly frustrated by my low regard for LLM safety. I didn’t get the role.
You raise good points and these would make sense under normal circumstances.
But none of this matters if AI offers infinite leverage and America hoards all the compute and talent.
This explains what I mean in a much better articulated fashion - https://x.com/tszzl/status/2065939227167392147
For other sovereign nations, being a market for American technology products is no longer good enough to get access to American tech. I imagine the relationship would become a lot more one-sided (in favor of America) if the model capability gap keeps increasing.
I humbly accept it's very difficult to game this out with any degree of confidence, especially as other countries deploy more resources.
But the questions about viability of the labs because of export restrictions is not in my cone of uncertainty. If you believe the labs' implied/stated objectives, the end goal is eating all human-driven GDP, and the US is still the largest single-market economy in the world, last I checked. Keeping the politics of AI-driven unemployment aside, economy-wide automation would make the US wealthy beyond imagination.
US exports as a whole is only 10% of the GDP. I am not aware of the international revenue share for OpenAI/Anthropic.
I'm assuming this is popular because of Fable restrictions. AFAIK, open source is not excluded from ITAR / EAR restrictions (or other export restriction in other countries).
So the real solution you're looking for is technology that can't be arbitrarily gatekept by a sovereign nation.
We saw some of this behavior in the last administration too. US voters deserve better.
With due respect, this take is very deluded. US voters have very little to lose if the tech is not available to the rest of the world. US politicians and elite, regardless of political inclination, understand the enormous strategic potential of this technology and will ITAR the shit out of frontier models and/or use them as leverage for extracting concessions out of other countries.
The main losers are Big AI labs, their investors, foreign employees and rest of the world.
Fwiw, China and other countries would’ve done the exact same thing. It’s perhaps the game theoretic optimal approach when your comparative advantage is so vast (capital, compute, talent, embedded knowledge) and keeps growing especially if RSI is real (making it nearly impossible for anyone else to catch up)
To the rest of the world, US aggression is perceived as oppressive and unilateral primarily because it’s a political act meant to further domestic political goals. All US administrations act in their own self interest, as they should.
But the US being the world’s cop was never going to be sustainable indefinitely. It was partly because of hegemonic power and partly because of a reluctance of wealthy democratic countries to seriously invest in their military when they could just “pay” someone else.
I don’t think anyone wants the pre world war 2 order. So we are likely going to see otherwise reluctant states step up militarily to protect their neighborhoods.
those are some massive actuators on those joints
Brakes are being added to actuators but they're more useful for static holds / locking than dynamic balancing. Standing even in humans is a dynamic balancing activity.
You evaluate the merits of the content? From a high-level systems pov, the article is largely correct, even if some details might be missed / simplified
I feel the same until I’m reminded I’m paying Anthropic $100 every month for something that’s indispensable to me now and would probably pay a lot more. Very inelastic demand as long as competition is low at the frontier.
3/3.1 Pro appears to have knowledge about eccentric topics with no obvious sources that often turns out to be right.
It does hallucinate a lot though, and is the most affected by context rot in multi-turn conversations
Acetaminophen (paracetamol) is the drug of first choice for addressing pain and fever, in India at least. To the extent that it's regularly abused, and I know people who have been hospitalized because of abuse.
Even then, doctors are usually disapproving of ibuprofen (or some combination of it with paracetamol) unless paracetamol is contraindicated for some reason, and I had always wondered why.
Yeah it's weird, almost like we're seeing two cults form in real-time.
I imagine there's a benign explanation too - the intelligence of these models is very spiky and I have found tasks were one model was hilariously better than the other within the same codebase. People are also more vocal when they have something to complain about.
In my general experience, Opus is more well-rounded, is an excellent debugger in complex / unfamiliar codebases. And Codex is an excellent coder.
They have several levers for demand destruction. From Anthropic's POV, I suspect this is least to worst bad
- reducing the surface area of "acceptable use" (e.g., blocking third-party tools OpenClaw)
- tighter usage limits and more subscription tiers
- increasing existing subscription prices
- moving to usage based model completely
- taking away compute from training next gen models (future demand destruction)
Every single one of them oversells their capacity
Indeed. And this model breaks in several cases that overlaps with the current AI business model:
- marginal cost of incremental usage is too high (Movie Pass)
- adverse selection (all you can eat monthly steak subscriptions)
- demand is synchronized (WeWork)
The only logical step for Anthropic now is to buy the Dwarkesh Patel podcast
I tried ls20 and it was surprisingly fun! Just from a game design POV, these are very well made.
Nit: I didn't see a final score of how many actions I took to complete 7 levels. Also didn't see a place to sign in to see the leaderboard (I did see the sign in prompt).
Gaming this out for peer adversaries is mostly moot, right? The post-Cold War strategic balance has mostly hung on MAD. And Russia, in particular, has responded to any attempt at building missile shields with more capable missiles.
It's likely more relevant for asymmetric conflicts that involve conventional weapons, and would enable an otherwise less resourced adversary to become a near peer.
Dennis Bushnell from NASA presented this deck in 2001, and is quite prescient about UAVs and distributed warfare.
https://alachuacounty.us/Depts/epd/EPAC/Future%20Strategic%2...
Reflected inertia does scale as the square of the gear ratio but it's a bit misleading unless you also consider the change in rotor inertia, which scales as a cube of the rotor radius (as the article points out).
The other side of the scaling laws say that motor torque scales as a square of air gap radius (roughly rotor radius), and output torque scales as linearly with gearing ratio.
When you balance these out, the reflected inertia depends on the inverse of power dissipated for a fixed output torque.
In an ideal world, your total reflected inertia is independent of the gearbox and largely depends on the motor fill factor and how hot you can run it.
I agree it is pretty awesome!
You would hit electrical steel saturation limits way before you need to pump in enough current to justify super-conductance.
Cooling in general is not a bad idea to allow you dissipate heat as you push motors to their saturation limits.
Very impressive. But it doesn’t solve the whole problem yet.
The robot and ball pose is estimated by high speed mocap cameras, and is fed to the policy.
I imagine estimating that with onboard cameras - how humans do it - is much harder.
Almost all of closed loop robotics is a state estimation problem. Control is “solved” if you can estimate state well enough.
I think ChatGPT has a huge advantage here. They have been collecting realistic multi-turn conversational data at a much larger scale. And generally their models appear to be more coherent with larger contexts for general purpose stuff.
It might be that this admin does not have the capacity to reason about second or third order effects.
But given that what would typically be red lines for previous administrations have been brazenly crossed without consequences, why would they bother?
In a broader context, both labs are engaging in "safety theater".
Neither know how to solve the alignment problem while market pressures are making them race towards capabilities (long horizon, continual learning) that will have disastrous consequences .
Wow. Surprising to see open hostilities between the leaders of the big ai labs. The differences appear to not just be competitive but also ideological.
Edit: Also openly calling OpenAI employees "gullible" and "twitter morons" seems sub-optimal if you like that talent to work for you at some point.
I think their consumer plans are gross margin positive but OpenAI has ~50M paying subscribers driving >$10B in revenue.
Realistically, you need at least ~1M subscribers to cancel to make this painful.
But I suspect this will get drowned out in the face of other news.