I've been using AI to write code recently. I figured I should give it a go.
Not reviewing and understanding and editing the output code is a recipe for disaster.

For example, in code for a cryptographic hash function there are padding rules to bring the data size to a multiple of the block size and add a length. So for example with SHA3, a minimum of 65 extra bits. If your data length mod the block size is 65 bits less than the block size, then add the pad bit, put the length at the end of the block and fill in between with zeroes.

If you are longer than that, then it all spills over into the next block and you add a block.

If your length exactly ends up fitting in the block size, then you add a whole extra block with just the padding bit and length and a bunch of zeroes.

The hash that claude spit out missed that last case, so creating a hard to find bug, where in one out of 512 bit sizes, the hash fails. If I had not spotted that by reviewing the code in detail, there would have been a catastrophic bug creating a security vulnerability and system failures in chips.

So read the code and fix it or be doomed.

The head of an AI company called "Amodel".

>These employees produced valuable work

Rocksmith 2014 is/was a wonderful thing. It's still on Steam and available with all the different DLCs.
Rocksmith's current version is crap subscription nonsense with greatly reduced functionality and none of the good music available in Rocksmith 2014.

Their last good output wrt Rocksmith was 12 years ago. I don't hold any hope of them returning Rocksmith to its former glory.

You are confusing speech with headlights. Headlights are the bright things on the front of cars. Speech is the nonsense that comes out of your mouth.

The keyboard could also be good. Like with mechanical keys and a satisfying thunk or thonk or click or whatever floats your boat. But I expect that it won't be that. The keyboard will be horrible.

>The statistical methods used medical studies are relatively much simpler than in say, engineering. That's not where gotchas are. So we need robust studies, a convergence of evidence, and meta-analyses from competent centers.

I don't follow (that last line). Engineering, at least in my area uses fairly simple models because we can get lots of data and we can control confounders because silicon doesn't care the same way that human subjects do. E.G. for PUF reliability you can measure the distribution of pairwise hamming distance across thousands of chips. This is more conservative than golden value hamming distance, but you compensate for the drop in sensitivity by getting more data.

The need for meta analyses in medicine comes about because of the large amount of underpowered studies in medicine and nutrition. The use of statistics in the source studies is often, well creative. The metas read like the statistical clean up crew. It feels like larger studies (costing more, I know) would lead to simpler statistics - but that's a guess.

Definitely I've come to 'trust' some researchers to do the research in a way that you know the claims match the data, because they've been consistent in doing that. Many others I just ignore for the opposite reason. Medicine has had its research problems, but nutrition research is where the real shit show has been going on for decades.

I'm involved in two companies after leaving Intel.
Everyone works remotely. There are no office costs to pay. We meet online.

If you're starting a company today and you want to be free to employ people for their skills rather than their location, this is the way to do it. Subscribe to one of the many online conferencing services and get to work.

>Unless you have at least a masters in the related subject, ideally a PhD

I disagree. You can spend a few years studying the nomenclature, norms, methods and underlying science and be entirely capable of reading and understanding a paper in context for a field that is not the one you started in.

Sometimes it's what you bring to the party that helps. I bring some knowledge on statistical inference and experimental methods, which arises from my day job. My interest was understanding my own health. It took about a decade of reading papers and textbooks to get up to speed. It has freed me from listening to health advise in media and not knowing how to tell if it's sound. I can go to the sources and see them in context.

If you want a difficult statistical environment, try education - That was my wife's PhD topic. My domain has no shortage of data. I can make all the data I need from silicon. The difficulty is in understanding it and what to do about it.

You could read my most recent paper ( https://dl.acm.org/doi/10.1007... ) and understand it with only a solid grasp of school level algebra and a spot of probability. I wouldn't expect anyone not involved in my field to care one bit about this algorithm, but I like it. It's neat.

Her comments on the nature of the threat from Russia and China are well put and stand up to analysis. That she was stating these things in public suggests that she wants the politicians to stop dithering and she is correct about that.

Her comments on tech seen naive. The tech world won't take her seriously and with good reason.

Yes. There is clearly more than one "cause", but the proximate cause is chemicals tricking the immune system into attacking specific cells.

What the MAHA people don't do is read papers. I do.

E.G. Here's a train of thought:

1) A long time ago, in the UK, a study showed a significant correlation between Parkinson's and exposure to insecticides.

2) More recently, a study showed lectins from wheat forming a ring around the vagus nerve and traveling up it to the Parkinson's site in the brain where biosimilarity between the lectins and tissues in the brain set up the autoimmune reaction that is part of Parkinson's. This wasn't a dodgy correlation study, they took photographs.

3) 99+% of the insecticides modern human's encounter are the "natural" insecticides in plants.

Conclusion : I'm suspicious of wheat and not for the usual reasons.

A long time ago I helped my wife with writing R scripts for her PhD. Why? MANOVER. Your average stats library will do ANOVA (Analysis of Variance) but not MANOVER (Multivariate Analysis of Variance). R did the MANOVER. It could also read in the CSV and had plenty of distribution models to apply.

Of course, I read up and wrote my own MANOVER implementation in python so I don't have to touch that horrible R language ever again. There's a little bit of R in my RNG book, but that's because it's in a section reviewing the distribution support in languages and so R is there because that's what it does well.

When I'm in charge, that's the font they'll be required to use.

Indeed it covers more than one area of law.
The two that were in my mind were (1) Copyright infringement and (2) patent infringement.
However I was using "theft" colloquially since I'm way too lazy to write like a lawyer.

When the AI steals the ideas of others and presents it as a new idea to the AI user, it's still theft and the inventor is the original inventor and the ideas were on the open internet to be scraped by the AI and so is prior art.