RAG to Riches
sourcegraph.com
sourcegraph.com
It's not something new and predates ChatGPT by a long shot.
Some 10 years ago I was trying to market my trading system and the backtesting/simulation software that produced it and there was one and only one question that I was asked by a VC I got in contact with: "Is it AI?". I made the mistake to say "well, no.. plain mathematics". Never heard of them afterwards.
Now I'm soon going to market a new trading system and this time you know what I'll say that it consists of ;)
Incidentally I did work with "AI guys" trying to somehow drain some money out of the stock market using some form of magic. The approach is very "Lords Of The Rings" like, you don't need to know fundamentals, don't need a keen understanding of the processes and various instruments, because "the ring" knows. All you need to know is to spin the ring around your finger and make the wish "make money".
I never saw anything but bullshit coming out of these "AI in finance" but at the same time never saw but "great successes" reported by these guys twisting and twirling and tormenting the numbers in desperation to make them look profitable.
The current craze is magnitudes larger, of course, but we have not had an AI winter for a long time.
What's the similarity to Lord of the Rings? That sounds more like Aladdin.
I recall (very badly) having read some Carl Sagan's reflection on this and how the fact is the majority of people hold some form of "magical thinking" which is completely at odds with rationality and the scientific method. Basically all you need is acquire some form of "magic ring" and you can be the "magician". When in fact you are not, you only are the magician if you can forge the magic ring. And diagnose and repair it when it malfunctions.
But that's not what "AI finance" is about, that's classical finance. AI is the magic ring here, somehow magically "it" thinks and does so that you don't have to. You only spin the ring and enjoy the riches.
The only thing the examples there have in common is that they're all rings.
They do different things; what makes "magic rings" a leitmotif when "magic hats" (also very common) aren't? The rule in fairy tales is very simple: any object can be magical.
You have magic swords, magic knives, magic bags, magic shoes, magic dolls, magic paintings, magic lakes, magic trees, magic stoves......
But this is not just about fairy tales, it’s also about folk lore. Things that people in real life sometimes actually believe in.
The linked Wikipedia article says:
> A finger ring is a convenient choice for a magic item: It is ornamental, distinctive and often unique, a commonly worn item, of a shape that is often endowed with mystical properties (circular), can carry an enchanted stone, and is usually worn on a finger, which can be easily pointed at a target.
Seems reasonable that in real life a ring might make a more convincing magical item than a hat.
People have been wearing special hats for symbolism for several thousand years.
And on top of all of that, the specific claim was that rings have a special presence in fairy tales, which is definitely not true.
In the 90’s you could prefix your company name with “e-“ or suffix it with “web”, or actually have the “.com” in the registered company name and your stock would immediately skyrocket.
(I’m not kidding or exaggerating - things got that OTT before the bubble burst).
https://en.m.wikipedia.org/wiki/Dot-com_bubble
“As a result of these factors, many investors were eager to invest, at any valuation, in any dot-com company, especially if it had one of the Internet-related prefixes or a ‘.com’ suffix in its name.”
https://www.bloomberg.com/news/articles/2017-10-27/what-s-in...
Is what these AI conversations sound like to me.
I'm a ML Engineer and for the past few months I'm getting as much activity in LinkedIn as at the cusp of 2022 job market. In the startup space, I see 2 major trends:
1. Startups with the mission to solve problem X using AI.
2. Startups that solve problem X, but now are pivoting to AI.
The funniest case was this startup that was hiring for their founding ML team, and the recruiter simply couldn't tell me what this team was supposed to do.
Also, in the next few years all the sauce and moat of these kinds of systems will be in the retrieval architecture. I wish they would go into further detail (just a little bit at then end).
In any case great read! Might even give it a try.
Disclaimer: I work for Sourcegraph.
Only those applicants who are from Korea.
Here is a summary with some additional context:
1. AI competition is heating up, with companies like OpenAI, Anthropic, Google, Meta, etc investing heavily and releasing new models. There is also a proliferation of AI startups raising large funding rounds. OpenAI fired then quickly rehired its CEO Sam Altman, causing confusion.
2. Machine learning engineers are in extremely high demand in the job market, receiving lavish recruiting offers. Meanwhile, other engineers like backend developers or mobile devs are struggling to move jobs since companies are mainly focused on security and ML hires. ML people get showered with "bear spray" level recruiting interest.
3. Most companies, even big Fortune 1000 brands, are still only in the initial stages of developing a concrete AI strategy and execution plans. There's a lot of "we're working on it" talk but not much to show yet. Companies are slow to focus on this due to inertia and competing internal priorities.
4. Metrics like "completion acceptance rate" (CAR) for coding assistants don't actually quantify developer productivity gains. However, some companies wrongly want to use CAR to closely monitor developers.
5. Cody, a new coding assistant from Sourcegraph using retrieval-augmented generation, leverages Sourcegraph's code search and understanding strengths. It is now generally available. Comments in code can improve Cody's understanding. Sourcegraph has developed Cody quickly thanks to its existing infrastructure.
Its marketing for a new AI assistant whatever. Waste of time.
Did you read the whole thing?
Every time i try to generate Rust code, all LLMs perform very good, while by trying to generate untyped code like Scheme, all of them perform okey-ish at best.
Better data can be produced for LLMs, like cargo-public-api to teach the machine about the api exposed and afterwards the code, as well as cargo-modules to teach the machine about the hierarchy of modules and only afterwards the code and so on.
Tricks like that, also mixture of experts architecture for LLMs to second think their response, will put an end to the coding soon. My point is, that coding assistants will not exist soon, because coding will be a thing of the past.
I can't find any source for this claim.
>Although Elon’s proprietary LLM is still held up in making sure it helpfully tells advertisers to go have marital relations with themselves
I can't find where Grok is setting up marriages between advertisers.
>Zuck chose not to execute the employee who leaked it.
LLAMA was leaked by someone who applied for the weights and not an employee.
>investors have still managed to raise roughly 2.3 quadrillion dollars to dump into AI startups.
This number is hallucinated.
>This resulted in approximately 23,342,971 news articles about how you should hide in your basement because they’d created Skynet
Another hallucinated number.
>and it was discovered that the board had accidentally fired a deepfake of him
Deepfakes are not related to this event. Hallucination due to being related to AI.
>Real Sam pointed back and said, famously, “No, YOU are the one who is fired!”
This quote is hallucinated.
I could keep going. Every few lines you will encounter something that is made up or wrong.
This is of course also true of every other standup comedy routine in history, but it was only a scandal for Hasan Minhaj.
There appears to be a breed of modern people who aren't capable of understanding the concept of humor. Unfortunately for Minhaj, those people were drawn to his act as an expression of piety, which isn't what it was.
This is classic Steve Yegge hyperbole. He's been ranting like this for decades[0].
His (accidentally public) rant comparing Google and Amazon business practices is a classic[1]. In it he says, "Jeff Bezos is an infamous micro-manager. He micro-manages every single pixel of Amazon’s retail site." Surely no one would take that literally.
[0]: https://sites.google.com/site/steveyegge2/home
[1]: https://courses.cs.washington.edu/courses/cse452/23wi/papers...
After downvoting the parent, it occurred to me that I was basically picking on a handicapped person. So I felt compelled to vote it up instead, to help redress the heartlessness of my fellow HN'ers as well as my own. "Be the light you want to see in the world," that sort of thing.
That doesn't mean my initial vote was wrong, though; it just means I felt bad for the author.
it couldn't be further from LLM style - not that LLMs don't write good text, but it's bland and has that unmistakable AI feel, almost always {intro, 3 paragraphs, conclusion} and making sure to name opposing views or warnings
Everyone who’s building rags has to meditate on this.
Completion Acceptance Rate may not be sufficient for measuring productivity but it does seem like a valuable measure of usefulness, although it can be useful due to accuracy at best but also laziness, carelessness, or ignorance. A manager who wants to reduce head count can interpret high-CAR programmers as among the lazy, careless or ignorant if so inclined.
Damn. You know this is going to happen.
lmao