I think LLMs are trained to not refactor. I think it’s either that you would need to do something in training to make them want to do it and the labs don’t do that, or that the labs correctly guess that it would be very annoying for LLMs to go and refactor your existing code as they go. This creates bad effects (eg crazy hacks to avoid refactoring and, much worse, not refactoring the code they only just wrote as required) but I think the alternative would be worse – it’s not something you always want to read and the refactoring is often done incorrectly, restructuring the code to the best shape for the current task rather than something that balances many different needs.
I think being successful makes hiring easier because you can source better candidates, and being big makes hiring easier because candidates are more likely to know people at the firm who can use referrals to work around otherwise broken systems.
It’s perhaps also worth noting that lots of companies used to copy how Microsoft did interviews and later they copied how google did interviews so clearly there were some ideas that those companies were good at hiring. (I’m not sure this strategy was that good. The problem for the Microsoft or google type companies is filtering out acceptable hires from a deluge of applicants with acceptably low errors and costs; the challenge for less desirable firms is sourcing candidates who are both high quality and not about to be hired by Microsoft or Google)
One company that comes to mind when I think about being good at hiring was one that recruited a bunch from my university around when I was graduating. Their particular specialty was hiring illegible graduates with a lot of potential (eg classicists, science students without little programming, etc), training them well, and effectively underpaying them a bit for how skilled they were (which only worked out because the UK has a pretty shit job market for tech and because those people liked working there). I think it was more effective for them than trying to hire the same computer science graduates as everyone else would have been.
I think various ‘longer interview’ processes can be good by reducing the chance of particularly regretted hires. This could be internships (but note this goes two ways and you want interns to accept offers and recommend the programme to their friends even if they are not hired) or work sample tests. Both have the downside that they are more work for the candidate (especially internships or some other short-term-to-possibly-long-term position) and so experienced candidates who feel they have better options and less need to prove themselves typically won’t take part (this depends a bit on how much they want to work at your specific company of course). Potentially this isn’t so bad – competing to hire the same people as everyone else is going to be more expensive – or potentially it is bad – maybe there’s a reason those candidates are in high demand and you will suffer from only getting a look at people who didn’t fit the typical pattern. I think it’s going to depend a bunch on how good you are at sourcing candidates and how hot your firm is.
Their IPO sees the company as doing a small amount of space and $22T of B2B services. If you believe that then space launches shouldn’t have much impact on their valuation compared to things that affect their hypothetical services revenue. If you think the value comes from the rockets, you would need some very large multiples to justify their desired valuation in which case, sure, the rockets are the cause of volatility…
The super literal interpretation ideas were much more common in the past when LLMs didn’t exist. Now we have models that are generally pretty good at picking up on nuance and understanding what you mean but also often quite bad at execution, which is roughly the opposite of that idea. I think reward hacking is perhaps the closest we see llms get to literal/malicious interpretations of instructions.
LLMs are neither of those. They're quite good at pretending they understand what you mean, but they don't. That's why they can't execute: they're mimicking the form, not the substance, and then we see the form and anthropomorphise them in our minds.
I've repeated the argument over and over since the GPT-2 days, when I derived it theoretically by inspecting the architecture of the model. I am now fatigued, and enough other people have taken up similar arguments – some developed half-way to a mathematical proof – that I no longer feel the obligation to keep repeating myself.
Firstly, consider why so many people like running outdoors when they could just run on a treadmill and watch tv indoors? Secondly, note that running injuries don’t tend to suddenly appear and cripple people half-way through a run. They are more likely to appear at the start, sometime afterwards, or to gradually progress in a way that might lead one to reducing volume or resting.
I actually don’t know the details of the specific crimes. Eg if you’re a soldier and you post on Facebook that you’re about to go on a raid to depose a head of state, that’s presumably a secrecy violation you would be punished severely for. The insider trading can be like this too in that you’re improperly using the information you are privy to due to your being an insider. If you’re a congressperson and you tweet that the government is about to do such a raid, I don’t know what the legality of that is – perhaps you have some kind of privilege to reveal these things and any censure must happen politically (eg impeachment, losing elections, etc) rather than legally. I don’t know what the rules for insider trading would then be – legislators are not insiders in the way that soldiers are.
Ignoring the moral argument, it isn’t all that clear to me that this would actually be a crime for a legislator under US securities law. It may be that new laws would be required to be able to punish legislators for this kind of behaviour.
He was charged with "unlawful use of confidential government information for personal gain, theft of non-public government information, commodities fraud, wire fraud, and making an unlawful monetary transaction.". Supposedly, unlawful use of government confidential information could also be applied to legislative and other people in the government
Congress sometimes includes an exemption for themselves from some crimes. Others are excused by the Constitution:
> The Senators and Representatives shall receive a Compensation for their Services, to be ascertained by Law, and paid out of the Treasury of the United States. They shall in all Cases, except Treason, Felony and Breach of the Peace, be privileged from Arrest during their Attendance at the Session of their respective Houses, and in going to and returning from the same; and for any Speech or Debate in either House, they shall not be questioned in any other Place.
> The law prohibits the use of non-public information for private profit, including insider trading, by members of Congress and other government employees.
I’m not sure if it’s AI so much as a composition of dozens of images stacked on top of each other. The shadows of different objects seem to be going in different directions.
reply