> They also bragged (for Fable) about how they "ran an external bug bounty that produced no universal jailbreaks in over 1,000 hours of testing" only for it be circumvented almost immediately.
Where did you see there was a universal jailbreak?
How do you weigh the DOD fight against warning about Mythos' dangers when determining what made Anthropic 'the tall poppy'?
"Response rates for this data collection were relatively low, both for the United States and for several other participating countries. There is evidence that procedures implemented to reduce bias associated with nonresponse have done so, and that the data are representative of the population. However, readers should be aware of the potential for bias and use caution when interpreting PIAAC results."
These stats don't pass the smell test. About a third of people in the US have a bachelor's degree, but only 13% can pass level 4/5 literacy challenge? If you dig into the sample questions, they are not hard. A level 4 task has the person read a short article and pull out the criticisms of some products.
I know not everyone with a bachelor's degree is 'smart' but it's hard to believe 2/3rds couldn't pass level 4/5.
Also 13% have a master's degree, does that mean those 13% are the only people passing level 4/5?
During COVID my company had mandatory days off (I think 14) if you reported any COVID symptoms. Those days were unpaid of course. The cherry on top is the people paid the lowest were the ones who couldn't work from home and were most likely to get COVID. This was pretty common at other places too.
> 100% of today’s SWE tasks are done by the models.
I do think he was overstating the current state of the models by a bit, but this is taken out of context. He is not saying this is where the models are at today.
He gives a spectrum [18:30] of the models taking over the SWE jobs:
- Model writes 90% of code (today)
- Model writes 100% of code
- Model does 90% of today's SWE tasks (end-to-end)
- Model does 100% of today's SWE tasks
- The SWE job creates new tasks that didn't exist before
- Model does the new SWE tasks as well (90% reduction in demand for SWE)
How do you parse the difference between marketing and having values? I have difficulty with that and I would love to understand how people can be confident one way or the other. In many instances, the marketing becomes so disconnected from actions that it's obvious. That hasn't happen with Anthropic for me.
I really hope Anthropic turns out to be one of the 'good guys', or at least a net positive.
It appears they trend in the right direction:
- Have not kissed the Ring.
- Oppose blocking AI regulation that other's support (e.g. They do not support banning state AI laws [2]).
- Committing to no ads.
- Willing to risk defense department contract over objections to use for lethal operations [1]
The things that are concerning:
- Palantir partnership (I'm unclear about what this actually is) [3]
- Have shifted stances as competition increased (e.g. seeking authoritarian investors [4])
It inevitable that they will have to compromise on values as competition increases and I struggle parsing the difference marketing and actually caring about values. If an organization cares about values, it's suboptimal not to highlight that at every point via marketing. The commitment to no ads is obviously good PR but if it comes from a place of values, it's a win-win.
I'm curious, how do others here think about Anthropic?
It's also the fun path! Although I think we should acknowledge that most don't have the means to do this. For engineers with a high salary I would advise saving as much as possible so you have more agency.
This is a point in time for the US and there are institutional paths to change. The comparisons to China forget that China does not have the same mechanisms for change. China is an immutable state outside of revolution or the administration just deciding to transfer power.
If they are successful in destroying democracy, I will reevaluate my view. We don't know what's going to happen in the midterms or 2028.