mgraczyk

36 ms·

631.
by mgraczyk 4y ago
I think there's a confusing narrative here, where people think layoffs is the same as not hiring or paying less. The market is still hot for roles that are competitive. It's still easy for many good engineers to switch jobs for a
632.
by mgraczyk 4y ago
Another important lesson is to look at Luna and understand the difference between it and tether. Luna was highly leveraged and had essentially no collateral with exogenous value. Tether is sketchy and potentially fraudulently under collater
633.
by mgraczyk 4y ago
Actually a good example, you should look at the actual losses incurred by investors in Madoff's fund. It would probably be educational if you think tether could "go to zero".
634.
by mgraczyk 4y ago
Maybe this is a good learning experience, and you could use it to update your beliefs about financial markets? If you're directionally and quantitatively wrong about some prediction for many years, it's very likely you are missing
635.
by mgraczyk 4y ago
It doesn't work though, that's the point. It's often difficult to convince people that something they have been doing for a long time doesn't work.
636.
by mgraczyk 4y ago
Beautiful! I don't have an iPhone, but just from watching the demo videos I am impressed by the design and UX. Good luck!
637.
by mgraczyk 4y ago
They aren't getting paid more, the companies simply make less money. Look at PPP wages over time
638.
by mgraczyk 4y ago
The part about Google still hiring talented people is not true. You should look around at offers. Google is hiring fewer good people because other companies are paying more. Google is losing their best people to startups that pay more. For
639.
by mgraczyk 4y ago
I went through the process of getting TS clearance the summer after my senior year of college, and I made the mistake of mentioning on my form that I drank alcohol with friends every weekend, and that I sometimes drank more than a few drink
640.
by mgraczyk 4y ago
Thanks for the response! Your point about Halide not being as predictable and having few guardrails is good. Yes I work on image processing, same pipeline as RAISR actually.
641.
by mgraczyk 4y ago
Looks cool. Are you familiar with Halide? In terms of target support or implementation complexity, how would you company using highway vs using Halide for similar projects? (I work at Google, use Halide)
642.
by mgraczyk 4y ago
I don't think so, and even if that were the case you've described the locations where a huge portion of SFs population lives. I currently live in Haight, it's nearly just as bad as those you listed. I'm on Divisidero rig
643.
by mgraczyk 4y ago
I've lived in SF for over 6 years, in at least 10 different neighborhoods from fidi to Mission to lower pac heights. This is completely accurate in my experience.
644.
by mgraczyk 4y ago
I read the entire thing lol. I don't agree with your conclusion. The paper doesn't make that conclusion.
645.
by mgraczyk 4y ago
Yes I boomeranged for a total of around 4 years, (ldap mgraczyk). I am aware there is a lot of third party code, but compared to the rest of the world, google is very averse to open source that it doesn't control. See for example front
646.
by mgraczyk 4y ago
Part of the joke here is that Google is borderline allergic to using open source software that it doesn't maintain.
647.
by mgraczyk 4y ago
Analyses like this are bizarre to me. There is an implicit assumption here that human generated content is often high quality and worth consuming or using. My experience, as an adult who grew up with the internet, is that close to 100% of t
648.
by mgraczyk 4y ago
No, I just don't believe the genetics database reduces my autonomy any more than going outside or posting a photo of myself online. You're attacking straw men here.
649.
by mgraczyk 4y ago
From the paper: > While most people would agree that putting a serial killer behind bars is an acceptable tradeoff, questions are nevertheless growing about who should control the genetic data compiled by genetic testing companies and ho
650.
by mgraczyk 4y ago
The comments here seem to misunderstand dynamics at big companies. Engineers often find themselves in meetings with 10s of people where there is no one person who could unilaterally reevaluate the usefulness of the meeting. Everyone in the
651.
by mgraczyk 4y ago
Which paper? The one pointing out that links are easy to determine with only a small genetics database? Why would I care about that, I want law enforcement to have (audited) access to 100% of the populations genetic records, so I'm not
652.
by mgraczyk 4y ago
It's implausible to me that the existence of a gene database would matter. Historically it hasn't been necessary to carry out mass killings. I don't think science fiction is a good source for designing policy.
653.
by mgraczyk 4y ago
I like saving lives with technology, even if we don't save all possible lives.
654.
by mgraczyk 4y ago
And if the probability of that happening AND the genome database being necessary for that government to carry out it's murders is smaller than 1e-7, then the database is good. I think it's much smaller than that, essentially zero.
655.
by mgraczyk 4y ago
On the other hand, solving actual crimes today has to be weighed against hypothetical future misuse by a hypothetical "future unfriendly regime". People have different beliefs about the probability of such a future, but to me it s
656.
by mgraczyk 4y ago
Most people who are into blockchain stuff would not use the word "decentralised" to describe a system where a municipal government has unilateral control. The thing that they solve well is incentivizing people to operate a network
657.
by mgraczyk 4y ago
I haven't been at facebook in a while, but seeing that chart at the top which shows the goal metric increase and become statistically significantly higher still makes me subconsciously very happy. One thing a lot of people outside the
658.
by mgraczyk 4y ago
I moved to Texas during the pandemic, primarily because I didn't want to live in a city with heavy lockdowns or with most people risk averse. I think it's great that different cities had different responses, but my preference was
659.
by mgraczyk 4y ago
It does select a function. The distribution over outputs conditioned on the input prompt is a different distribution, hence a different (random) function. It's not important that there are no weight updates. The new function, the model
660.
by mgraczyk 4y ago
I disagree with what you're saying here. > The input prompts to GPT-3 and friends are not "training examples". Training examples are labelled instances of a concept to be learned- that's PAC-Learning. LLM prompts are

More ›