Whenever I you tried to develop using guardrails with LLMs, I found out that they are much better at ,,cheating'' than a human: getting around the guardrails by creating the ugliest hacks around them.
8,476 karma · joined September 7, 2018
Whenever I you tried to develop using guardrails with LLMs, I found out that they are much better at ,,cheating'' than a human: getting around the guardrails by creating the ugliest hacks around them.
There are ways to report if a man has been sexual with a woman, but they somehow just don't get kicked out of the driver network.
Also just a simple example: Uber engineering blog is full of examples of how they rewrote their app in native Android then web then native again, but nothing about how to solve the real problems humans experience when driving with them.
It just feels that they view Uber as a simple logistic problem where drivers / riders are interchangeable and less like Tinder that tries to match people with similar scores abd kicks out the worst.
I remember adding lifetimes in some structs and then wanted to use generics and self pointing with lifetimes because that made sense, and then it didn't work because the composition of some features was not yet part of Rust.
Another thing: there are annotations for lifetimes in function signatures, but not inside the functions where there is a lot of magic happening that makes understanding them and working with them really hard: after finally the borrow checking gave me errors, that's when I just started to getting lots of lifetime errors, which were not shown before.
Rust should add these features but take out the old ones with guaranteed automatic update path.
The best engineer I know who can work with huge complex systems in a big company usually starts with a complex solution then after he understands what he wants to achieve thinks backwards and reimplements it in the fewest possible lines of code change with the already complex system.
For coding I love codex-5.3-xhigh, but for non-coding prompts I still far prefer o3 even if it's considered a legacy model.
I can imagine that its higher tool use is too expensive to serve, but as a pro user I would love it to come back.
They can just improve on it a lot.
if OpenClaw usage go up, and a service (OpenAI it looks like) gets lots of usage data for personal assistent usage, they can optimize to make it better for people who get a $200 subscription just because of that use case.
Elon Musk already denied Starlink for being used for remote killing, but at some point all these technologies will be nationalized, as they are too important not to be.
The only question is whether the safety of the models were really done well enough to protect the people and be a net positive force in the world.
I guess if they would be safely trained to do more good than bad (how Dario and SamA said), there wouldn't even be a need for the contract terms.
And if UK is precedent based, how come the previous precedents don't apply here?
I agree that no toilet breaks is cruel, but the problem here is knowing about the supplier using it?
There was not much about the legal bases in the article.
I have a different view point on what to automate and I'm working differently with agents, but I much prefer seeing projects like this on HN to just product announcements.
Why would killbots be sensible moderate with the number of hallucinations LLMs have right now?
They just need to have one rm -rf bug somewhere to so something disasterous, and at least Antrhopic's CEO understands the limitations of the software.
Deescalation stopped because of people in general not caring enough (and making money of being the biggest power), not because of administrations that come and go.
As to the immigration situation: we know that governments are not executing in general how they should be, but people are able to enforce some policies if they fight together united and in agreement. But right now they are not in agreement.
Maybe people don't agree with ,,nuke them'', but OK with USA starting nuclear experiments again (which USA is preparing for right bow), which is a clear escalation.
Russia is waiting for USA to start the nuclear experiments to start them itself for defending itself to be able to do a counterstrike if needed.
After that there will be no stopping of Japan, South Korea and Iran rightfully wanting to have their own nukes.
You don't have to have the ,,nuke them'' thinking, even one step of escalation is enough to get to a disastrous position.
People in the world have limited experience about war.
We're living in a world where doing terrible things with 1000 people with photo/video documentation can get more attention then a million people dying, and the response is still not do whatever it takes so that people don't die.
And now we are at a situation where nuclear escalation has already started (New START was not extended).
It would have been the biggest and most concerning news 80 years ago, but not anymore.
I feel that Sam is saying what investors want to hear, but the coding work it is capable of and how it improved with using the terminal (TerminalBench) in such a short time is something that I'm sure can't be seen by short term revenue projections. I'm sure the other AI companies are having the same speedups, but it's real.
The usual limit is of course the slop output that is not well modularized that makes it hard to do bigger things, and codex is terrible at refactoring into the right direction (it has no taste).
3x YoY growth in revenue is just not hard to imagine with this kinds of models, I think they have to get out with more expensive parallel working agents and higher-than-pro subscriptions, but it is coming I'm sure.
With Robotaxi it will get even higher as it will be clear 100% the company's fault.
But capital structures and politicians are still too close to old European companies from the second world war and don't allow venture capital to florish.
It's easier to earn money by winning a fake EU tender and giving back half of the money to a politician than doing something innovative.
Also in the USA you just wait 4 (or 8) years and you have a new president. In many other countries you don't have that luxury.
In software San Francisco is still the top for AI research: even when Peter Steinberger didn't know what he will do with OpenClaw, it was clear to him that the only place to move to was USA.
Terrence Tao was a good example of what happens when an exceptionally smart person stops getting funded by an American University: not moving to another country, but got VC money and created a new company.
USA politics is looked at so closely, because it matters and changes and still more democratic than most countries in the world even though democracy is a mess (as it's supposed to be).
For me it looks like they will reach parity at about the same time, so camera only is not totally stupid. What's stupid is forcing robotaxi on the road before the technology is ready.
Technology is just not there yet, and Elon is impatient.
Training pipelines are full of data preparation that are first written on CPU then moving to GPU and always thinking of what to keep on CPU and what to put on GPU, when is it worth to create a tensor, or should it be tiling instead. I guess your company is betting on solving problems like this (and async-await is needed for serving inference requests directly on the GPU for example).
My question is a little bit different: how do you want to handle the SIMD question: should a rust function be running on the warp as a machine with 32 long arrays as data types, or always ,,hope'' for autovectorization to work (especially with Rust's iter library helpers).
This is stupid. All you need to do is to get used to being hungry to the point of losing 0.1kg/day and measure yourself a lot (I'm doing the same).
Actually for me working out increases my appetite, and I feel like I have to eat so thar the gym session doesn't go wasted.
On the otherhand lots of people here are even more uncomfortable of the other option, which is quite possible: AI software algorithms may scale better than the capacity of companies that make the hardware. Personally I think hardware is the harder to scale from the two and this is just the beginning.