Is General Artificial Intelligence a Breakthrough or just a Financial Hype?

Mobile, so I can't write much. Theoretically anything can be vulnerable (especially when there are ultimately humans involved somewhere in a transaction or workflow or process. But as mentioned, the fact is banks trust cloud with trillions and they do process data within the cloud, not merely store data. For practical purposes, the threat today is not so severe to prevent such massive volumes of commerce, with any decent solution. Same with governments using the cloud. Security will have to evolve as threats evolve. The biggest threat, well beyond limitations of today's security capabilities with any reasonably architected solution, may well be self-extinction of their own enterprise if businesses do not consider using the cloud. But this feels like a 2000's conversation, businesses have migrated (well, apart from particular industries/verticals that are traditionally slow-moving or have complex legacy systems; they may have only partially migrated), it's been a couple of decades. 

I agree that "most orgs consider encryption enough of a solved problem", but that doesn't mean it is. If you are only storing in the cloud, and you encrypt before uploading using post-quantum cryptography, you are right, that should be safe enough for some time. However, once you are using the cloud to actually perform operations on your data, it decrypts it, leaving it vulnerable both to the insider threat of employees of the cloud provider, but also side channel attacks, etc. that potentially let others get access to your data while decrypted. (Not to mention that the keys had to be in the cloud as well). Homomorphic encryption allows processing the data *while it is encrypted*, thus even the server has no idea what functions it is performing, what the input or output really are, etc. Only you, the data owner, get the result and then decrypt it on your end (not in the cloud). Or have a "private cloud" which really means your org owns and controls the servers too so can implement additional security mechanisms (like only allowing your own software to run on them, have control over vetting employees who can access the servers, etc.)

Unfortunately, to date, while there have been some proof of concept experiments (keeping in mind the efficiency of known homomorphic algorithms is abysmal), it's still considered an impractical approach which is why I don't think cloud is practical for any non-public data processing until appropriate algorithms are discovered. Meanwhile, one can process on the edge, upload sanitized summaries if needed for sharing, and otherwise leverage cloud only for highly encrypted backups. 

I don't quite understand the point you are trying to make.

My opinion is biased by the fact that I've got many years under the belt, so manual coding has become a bit mundane. AI certainly helps me on that front. I also have the experience to direct AI to go on certain paths of choice and identify issues. A junior software engineer on the other hand, may well have exceptional knowledge of the coding language but IMO lacks the broader experience to see errors. So, IMO, they will struggle to get the same benefit from AI as I do, no matter if they are an insider or not.

Hence I see this as a huge breakthrough.

When you go into complot mode, please link to the source.

HKPhysicist said:

You should be able to find this video on Youtube or some paid video channels

Not convincing for a big claim.

To view recent AI tech from the insider viewpoint, please watch the dialogue between a reporter and an ex-X-AI top engineering/programmer.

The man, who "carelessly" disclosed certain hidden agenda and tie between X-AI and Tesla, claimed that they, as X-AI software programmers, were requested to start programming from the beginning definition.  They were not allowed to use AI assistant to code.

When they interviewed new potential employees with a certain software task, if the participant tried to code with the help of AI, this participant would then be filtered away.

I forgot his name, a recently fired ex-X-AI software programmer.  He was fired because he leaked certain secret strategical tie between X-AI and Tesla.

You should be able to find this video on Youtube or some paid video channels which tend to keep erased video clips.

Fair, but the decline in human code quality is more related to the abuse of bloated frameworks, libraries and languages with higher level abstractions, and the constant prioritization of "delivering fast" over creating performant and optimized apps (a priority imposed by most tech companies). The code itself is not that much worse, but there's a clear over-reliance in just shoving down "magic" dependencies into a project without caring about how efficient, lean, or optimized they are. With this they can get an app done fast, at the cost of being slow as hell, being a million MBs in size, or using a truckload of RAM (and sometimes all of the above).

The metrics used to measure the decline in quality due to use of AI however are slightly different; They have measured the abundance of duplicated blocks of code, the complexity of the code created even for the simplest tasks, the prevalence of "copy and paste" sections that have a bunch of unnecessary stuff, etc. All of which are a lot worse now with the help of AI than they were before.

battlecoder said:

All seem to agree that while the speed at which programmers are writing code has increased with the help of AI, the quality is decreasing fast.

I think we've seen a decrease in quality with human programming as well. Nothing beats the good old days, when it was important for your code to be written such that where the heads on the drum will be when the next instruction needs to be fetched, will contain that instruction, after all.

So we may just be seeing a sped up time-lapse movie of where programming is headed regardless. 256K words (a moby) of memory ought to be enough for anyone.

That's another problem with generative AI; The output is derivative. And once the input becomes the output of another model, you start seeing a steady decline in quality. It's a lossy process. Like compressing a JPEG, and then compressing it again, and again, and again. Eventually the image would be a mess of blurry blobs and nothing like the original picture.

Part of this has been shown to happen already in the world of programming with code-oriented LLMs; Several studies were published last year like this one,  this one or this one. All seem to agree that while the speed at which programmers are writing code has increased with the help of AI, the quality is decreasing fast.

What can be seen as a problem of entropy, is another major problem with that technology. Most people assume that AI "can only get better", when in reality, as human input gets diluted in a sea of AI-generated training data, it may actually get worse and worse.

One time, I asked Mistral to generate a simple Arduino code snippet for me.

Then, it showed me the codes and asked me to add a header file.

Later, I had found that the header file did not exist at all.

Asking programmer friends or buying a relevant good book is better. 

Hello all Friends who have joined the discussion.

Here is another viewpoint published on Nature about recent AI and LLM development:

AI models collapse when trained on recursively generated data